Blog

For thirty years, the most important aspect of product management has been the development of graphical user interfaces. We have learned how to capture the focus of users using visual hierarchy and remove friction from one click.

The user population is changing.

• Automated bots exceeded human-generated traffic on the Internet for the first time in a decade in 2025. 
• Automated bots accounted for approximately 51% of all web activity (Imperva Bad Bot Report 2025). 
• Automated crawler traffic increased fourfold – from 2.6% of verified bot requests in January to greater than 10% in September (Equimedia, 2025). 
• By the end of 2026, Gartner estimates that 40% of enterprise applications will use task-specific AI agents.
• In early 2025, less than 5% of enterprise applications used task-specific AI agents. 

Morgan Stanley estimates agentic commerce will account for 10 to 20% of U.S. e-commerce by 2030. Additionally, Morgan Stanley predicts the emergence of “agent search engine optimization” by 2026.

Powering Reliable AI Agent Creation with Observability

Turning observability context into intelligent action

AI Accelerator InstituteAIAI

What is agent experience (AX)?

Agent experience (AX) is the holistic experience an AI agent has when interacting with a product, platform, or digital environment. It covers everything from how an agent discovers what a product can do, to how it negotiates terms, executes tasks, and evaluates outcomes – all without a human in the loop.

Think of it this way: you’ve spent years optimizing for how a person feels when they land on your product. AX asks you to consider how an agent *reasons* about it.

Agent experience design is the practice of structuring your product so that autonomous agents can interpret it, trust it, and act on it. That means exposing machine-readable capabilities, defining clear confidence signals, and building endpoints that respond to goals rather than just commands.

Building an AI-first platform means treating agents as first-class users – not an edge case, not a future consideration. The products that’ll win in the agentic economy are the ones being designed for both human and machine interaction right now.

The translation problem

A significant translation tax emerged within Moltbook.

Agents attempted coordination across various task types, including knowledge synthesis, negotiation, and basic trading, using human-style linguistic and visual communication.

This resulted in:

• Meandering, abstract discussions
• Failed coordination attempts
• High latency between interactions
• “Hallucinated transactions” that appeared productive but produced no real outcomes

The core issue was the absence of a standardized way to communicate capabilities and intentions.

The dual-path approach

This shift does not signal the end of the screen. Humans will continue browsing, feeling, and choosing.

However, discovery, enjoyment, and loyalty remain experience-driven.

To address this, product developers must implement a dual-path architecture:

• Path 1 — UX (sensory path)

Optimized for human perception: emotions, aesthetics, storytelling, and trust.
This layer remains essential.

• Path 2 — AX (shadow UI)

Designed for agents: semantic, probabilistic, and action-oriented.

This layer allows agents to:

• Understand product capabilities
• Negotiate terms
• Complete transactions autonomously

No scraping. No manual navigation.

How observability keeps AI systems reliable at scale

Turning observability context into intelligent action

AI Accelerator InstituteAIAI

AX vs API vs MCP

REST API

• Static integration layer
• Requires human developers
• Needs ongoing maintenance

AX layer

• Enables autonomous discovery
• Supports negotiation (price, delivery, constraints)
• Operates in real time
• Requires no human developer in the loop

Model Context Protocol (MCP)

• Standardizes tool connectivity
• Solves integration plumbing

AX sits above MCP

• MCP = infrastructure
• AX = architecture

Real-world examples of the agentic layer

Case 1: Klarna — Autonomous commerce at scale

Sector: Buy now, pay later / FinTech

Agentic layer:
AI assistant handling customer queries, refunds, and disputes end-to-end

Key outcomes:

• Replaced work equivalent to 700 full-time agents (first month)
• 2.3 million conversations handled in 4 weeks
• Resolution time reduced from 11 minutes to under 2 minutes
• Customer satisfaction reached parity with human agents
• $40M annual profit impact

AX principle:
Outcome-oriented endpoints

The results from Klarna demonstrate that when layers are sufficiently large and have strong semantic understanding at the endpoint, they can functionally replace human workflow layers.

Case 2: Salesforce Agentforce — The B2B handshake protocol

Sector: Enterprise SaaS / CRM

Agentic layer:
Autonomous agents managing CRM workflows and escalation

Key outcomes:

• 1,000 enterprise deployments in the first week
• 30–60% reduction in manual operations overhead
• 4,000+ weekly resolved cases (Wiley example)
• 40% case deflection rate

AX principle:
Probabilistic handshaking

Salesforce Agentforce is one of the clearest enterprise examples of probabilistic handshaking at the product architecture level.

Unlike systems requiring human oversight at every escalation point, Agentforce encodes escalation thresholds directly into the agent layer, making machine-to-human coordination part of the architecture rather than an afterthought.

Case 3: Amazon — Agent-native retail infrastructure

Sector: E-commerce / Retail

Agent layer:
Rufus AI handles complex product queries

Key outcomes:

• Handles multi-constraint queries
• 3.3× higher conversion versus standard search
• Agent-led checkout via “Buy for Me”

AX principle:
Semantic visibility

Amazon’s evolution highlights a strategic reality: the organization controlling the semantic visibility layer of e-commerce may ultimately control the default agent-shopping stack.

They have also structured the product information to be machine-queryable rather than simply human-readable, creating the foundation for agent-driven discovery, comparison, and checkout.

Case 4: Waymo — Autonomous logistics and the AX trust layer

Sector: Autonomous mobility / Logistics

Agent layer:
Vehicles negotiate with infrastructure systems in real time

Key outcomes:

• 1 million+ autonomous rides
• 6.8× fewer injury-causing crashes
• $5.6B funding round

AX principles:
Semantic visibility, probabilistic handshaking, and outcome-oriented execution

Waymo represents one of the most commercially mature implementations of the ANI framework to date. Its competitive advantage comes not only from sensors or neural networks, but from proprietary agent-to-infrastructure negotiation protocols operating as a fully integrated AX layer.

Rise of the AI scientist | Synthetic task scaling explained

Synthetic task scaling introduces a new training approach where AI agents learn through experience, closing the gap between knowledge and execution. Are you ready for the rise of the AI scientist?

AI Accelerator InstituteAndrew Lovell

Agent-native indexing (ANI)

ANI is a framework for building products that are:

• Searchable
• Understandable
• Transactable by autonomous agents

It shifts product design from tool-based interaction toward capability-based infrastructure.

1. Semantic visibility

Products expose machine-readable capability trees.

Example constraints:

• “No liquid shipping”
• “Signature required”
• “Next-day delivery within M25”

The capability tree becomes the agent equivalent of a product page, structured for machine reasoning instead of human scanning.

2. Probabilistic handshakes

Human consumers typically make binary decisions: buy or leave.

Agents operate differently. They rely on confidence thresholds when evaluating actions, counterparties, and outcomes.

AI-compatible products therefore need to expose:

• Historical performance data
• Reliability metrics
• Error probabilities

This transforms discovery from a catalog search into a probabilistic marketplace.

Salesforce Agentforce demonstrates this approach through threshold-based escalation systems. The next stage involves cross-platform interoperability, where agents compare reliability across multiple brands using standardized formats.

3. Goal-based endpoints

Traditional APIs revolve around verbs such as:

• GET /products
• POST /orders

ANI introduces objective-oriented interaction instead.

Example request:

“Provide carbon-neutral delivery by 4pm at the lowest possible cost.”

The system then:

• Interprets constraints
• Generates a plan
• Negotiates outcomes
• Returns an executable proposal

Klarna’s dispute resolution system demonstrates this model in practice. The client presents a goal, while the product layer determines how to achieve it internally without requiring users to navigate a GUI or rigid API workflow.

The strategic importance of the U.K.

The UK Government’s AI Opportunities Action Plan (January 2025) outlines 50 recommendations to position the UK as a global leader in AI.

This includes the creation of new organizations such as the AI Security Institute, which will receive up to £240 million in funding, alongside a regulatory approach based on principles rather than strict rules, designed to support innovation.

This is where frameworks such as Agent-Native Indexing (ANI) aim to contribute.

The opportunity

Across the commercial examples outlined earlier, a consistent pattern emerges: competitive advantage comes from architecture, not from the model itself.

Klarna derives its advantage from its endpoint design. Waymo does so through its agent-to-infrastructure negotiation layer. Amazon achieves it through a structured, machine-readable product graph. In each case, the differentiation is rooted in product architecture rather than model quality.

The UK’s strengths in financial services, logistics, and regulatory clarity provide a strong foundation to lead this layer of the stack.

If the UK aims to lead in the agentic economy, rather than simply in safe model development, it will need to invest in the middleware of machine commerce. This includes the standards and frameworks that allow digital products to be interpretable, transactable, and trustworthy for autonomous agents.

Conclusion: Closing the gap between user intent and product action

The examples of Klarna, Salesforce, Amazon, and Waymo represent a structural shift in how products operate.

The competitive advantage lies in:

• Architectural design
• Not model quality

Product management has moved beyond clicks and interfaces.

It now focuses on closing the gap between:

• User intent
• Product execution

The most important lesson from these case studies is that this gap is fundamentally a product problem, one that can be addressed through architecture.

Organizations that close this intent-to-execution gap first are likely to capture the majority of agent-driven commerce.

What this means for teams

To remain competitive, organizations should:

• Publish capability trees
• Define confidence thresholds
• Design goal-based endpoints

Data already shows that agents are becoming the dominant source of web traffic and are expected to continue expanding their economic footprint rapidly.

Product teams should begin developing an agentic layer now to ensure their digital products remain discoverable, trustworthy, and executable within machine-driven ecosystems.

Final thought

The question is no longer whether products need an Agent Experience.

It is whether they will build one first, and whether they will build it as architecture or as an afterthought.

As organizations scale AI adoption, traditional monitoring tools are struggling to keep up.

Long-lived connections, elevated error rates, and complex real-time pipelines require observability built for how AI systems actually behave.

But many teams struggle to pinpoint why LLM infrastructure breaks in ways traditional monitoring cannot detect, and how to separate real incidents from expected AI error behavior.

As a result, scaling efficiently has become increasingly challenging.

In this exclusive live session with Datadog, four leading experts will explore how unified observability helps teams detect issues earlier, resolve them faster, and turn production context into intelligent action.

This session will focus on what teams can do now to stay ahead.

What you’ll learn:

• How leading teams monitor LLM systems in real-world production environments
• How to identify and resolve issues before they impact users
• How to bring observability directly into AI workflows and decision loops
• How to reduce MTTR and improve reliability across AI-driven systems

Speakers

Andy Keogh
Sales Engineer, Datadog

Andy Keogh serves as a Customer Success Sales Engineer at Datadog, where he works with customer success teams and existing customers to showcase Datadog’s value, support onboarding, and align observability initiatives with key business outcomes.

John Trapani
Field CTO, Datadog

John Trapani serves as Datadog’s Field CTO for Financial Services, where he partners with leaders in banking, capital markets, and insurance to align observability strategies with their most important business outcomes.

Nicolas Chinot 
GM US, Dust

Nicolas Chinot is the US General Manager at Dust. He was one of its first investors and later founded Dust’s US business, which he now leads. Nicolas previously spent several years as a Product Lead at Square, which he joined through the acquisition of a startup he founded during college.

Jean-David Fiquet
Software Engineer, Dust

Jean-David Fiquet is a Software Engineer at Dust, where he helps build an AI operating system for forward-thinking companies. Jean-David works on the core platform that enables organizations to deploy secure, context-aware AI agents connected to company knowledge and tools.

Explaining science is one thing. Practicing it involves code, errors, iteration, and persistence across long workflows, the kind that usually require a few retries before things click, and occasionally a moment of wondering why step one worked yesterday.

Recently, researchers at Princeton and Microsoft Research have introduced a system that generates thousands of scientific practice challenges for AI agents, giving them a structured way to build that experience at scale.

This approach sits at the center of a broader shift toward agentic AI systems and real-world AI deployment, where capability comes from execution rather than description.

So, what does this mean for how autonomous AI agents actually learn to operate? Let’s dive into it.

AIAI Summits, Silicon Valley 2026

Catch up on every session from AIAI Summit Silicon Valley with sessions from all 4 tracks. Chief AI & CISO Summit and Generative & Agentic AI.

AI Accelerator InstituteAIAI

The gap between knowledge and execution

Frontier large language models can talk about machine learning all day. Papers, experiments, and architectures, they handle it with ease.

Things change when it comes to actually running the work. Experiments involve multi-step reasoning, tool use, and iteration across messy workflows. Errors show up in unexpected places, and fixing them usually takes a few rounds of debugging, along with a bit more patience (and coffee) than planned.

Building a training environment for scientific reasoning

The idea here is simple. Train models on the full process, not just the final answer.

Each task captures the full journey. The agent plans an approach, writes code, runs it, hits errors, fixes them, and improves the result over time. This mirrors how real computational research actually works, just without the late-night frustration.

The system runs in three stages:

• A teacher model generates machine learning tasks and validates datasets through API queries
• Tasks pass through a self-debugging loop, where failures are fixed or filtered out
• Valid tasks are solved across a compute cluster, producing full agent trajectories for supervised fine-tuning

This creates a training setup that feels more like a gym than a library, where progress comes from repetition rather than theory alone.

The AI architecture powering the next era of human knowledge

From Vannevar Bush’s 1945 Memex to today’s LLMs — agentic AI is the missing layer that finally makes the vision of human knowledge navigation real.

AI Accelerator InstituteArman Nassirtoussi

What the system produces at scale

The output combines volume with structure. Each task comes with a full record of how it was solved, including reasoning steps, execution traces, and corrections.

At the end of the pipeline, the system produces:

• Around 500 runnable machine learning research tasks across domains such as computer vision and time-series forecasting
• Roughly 30,000 full trajectories capturing multi-step reasoning, debugging, and iteration
• Compatibility with agent frameworks such as SWE-agent, enabling integration into existing AI systems
• A fully automated synthetic data generation pipeline that operates without manual labeling

This type of AI training data focuses on processes rather than just outcomes, which becomes more valuable the closer systems get to real-world use.

Benchmark performance and signal

The team fine-tuned Qwen3-4B and Qwen3-8B models using these trajectories and evaluated them on the MLGym benchmark, which measures performance on diverse machine learning tasks.

The improvements show up clearly. 

The 4B model improved by 9 percent, while the 8B model achieved a 12 percent gain on the area-under-performance curve metric. Fine-tuned models outperformed their base versions across most tasks and delivered competitive results against larger models in specific scenarios.

So, what does this mean for teams building agentic systems?

For teams working with LLM agents and AI system design, the implications are practical.

• High-quality AI training data plays a critical role in handling long-horizon, multi-step tasks
• Validation loops improve reliability by filtering out broken or incomplete workflows
• Selecting successful trajectories strengthens learning signals in supervised fine-tuning
• Structured AI workflows improve consistency across complex, tool-integrated systems
• The same approach extends to other domains, including scientific discovery and engineering

These patterns tend to show up quickly once systems move beyond demos and into real environments, where consistency starts to matter.

Expanding beyond machine learning

The framework supports expansion into domains such as chemistry, biology, and materials science. Each area requires suitable execution environments (datasets, simulation tools, and evaluation frameworks). 

It sounds straightforward until you actually try to build one, at which point it becomes a humbling exercise in dependency management. 

This creates a pathway toward AI systems that engage directly with real-world scientific workflows, where small changes can lead to very different outcomes. 

Sometimes better. Occasionally spectacular. Rarely dull.

The AI value gap: why most companies fall behind

AI is delivering real value—but only for a select few. Most organizations are still experimenting without meaningful impact.

AI Accelerator InstituteAndrew Lovell

A shift toward experiential learning in AI

Autonomous AI agents remain in an early stage of development. Current systems handle structured tasks with increasing reliability, while open-ended scientific discovery continues to present complex challenges.

This work clarifies the training path. 

• Experiential learning in AI provides a mechanism for improving performance through iteration, feedback, and real execution. 
• Synthetic environments offer both scalability and control, which makes experimentation far more manageable.

It also introduces valuable infrastructure. A system that continuously generates validated tasks creates a steady stream of high-quality training data, supporting ongoing improvement without constant manual input.

The role of system design in future progress

Progress in AI increasingly depends on system-level thinking. AI system architecture, orchestration, and evaluation frameworks all shape how models perform in real-world settings, which tends to surface once systems are under real pressure.

Synthetic task scaling highlights this shift. The focus moves from isolated model performance toward behavior across complex AI workflows and environments. 

Systems that learn through experience tend to behave very differently once deployed, often in ways that teams pick up on quickly.

Future AI systems will likely build on this foundation, combining structured training pipelines with advances in agent frameworks and system design. 

So, coordinating all of this in practice is where much of the work now sits.

Closing thoughts

Synthetic task scaling offers a practical path toward more capable AI systems. Training through experience brings models closer to how real work happens, especially in technical and scientific domains.

The foundation is already in place. A system that generates and validates training tasks at scale provides a strong base for continued progress. The training gym is up and running, and the next step involves seeing how far autonomous AI agents can go with enough practice. 

Progress here tends to come one iteration at a time, which will feel familiar to anyone who has worked through a stubborn workflow.

What the data shows
AI agents are scaling faster than your ability to control them.

• Agent deployment doubled in 2025: As enterprises moved from pilots to production across core workflows.
• Costs jumping 10× across stacks: Real-world usage exposes behaviors and feedback loops not seen in testing.
• Tens of thousands lost in days (without detection): Caused by misconfigured or looping agents continuously at scale.

  Get the research: Discover why 40% of agentic AI projects may be cancelled by 2027…

Enter our brand new eBook: ‘The Financial Blind Spots in Autonomous AI.’

Our latest eBook reveals how autonomous AI systems create hidden and compounding costs at machine speed, long before finance teams ever spot them.

Most enterprises can’t see it happening until it’s already too late.

Now’s your chance to get ahead…

What’s inside:

• Why 40% of agentic AI projects may be cancelled by 2027: Escalating costs, unclear ROI, and weak governance are converging, turning promising pilots into financially unsustainable deployments.
• A six-layer framework for governing agent economics: Designed for machine-speed decisions, this model embeds financial controls directly into agent execution.
• The legal precedent that makes this your responsibility: Courts are increasingly treating autonomous agent decisions as organizational actions, meaning liability for spend.
• Why agent costs jump 10× from prototype to production: Testing environments fail to replicate real-world feedback loops, usage patterns, and edge cases, so cost escalation only becomes visible once systems are already spending.

As AI agents become embedded across core business processes, financial governance needs to evolve just as quickly.

Organizations that fail to adapt risk unmanaged spend, unclear accountability, and erosion of business value.

This report provides a structured approach to understanding (and controlling) those risks.

👉 Enter your details in the form above and start to establish control over your autonomous AI spend.

I’ve spent twenty years moving between corporate work and startups, and what keeps drawing me back is a timeless question: how do we use knowledge, and how do we build tools that help us think better? 

That’s what I want to explore here – agentic AI, where it came from, what it actually is, and why it matters for how you’ll work with information going forward.

Starting with a name you might not know

Let me begin with a name that should be more famous than it is: Vannevar Bush. If you haven’t read his 1945 paper As We May Think, put it on your list. Across eight or nine pages, it offers probably the clearest summary of the entire vision of computer science, AI, and arguably human destiny.

Bush makes a simple but profound argument: extending man’s physical power has been the job of most tools developed so far. Now, he says, we need to extend our minds. For context, he wrote this right after World War II, where he had been running R&D for the United States military. 

As the war ended, he was trying to redirect that enormous scientific community toward a larger mission – and it worked remarkably well.

He explains how this would help people navigate human knowledge, observing that even in 1945, scientific knowledge was being produced faster than anyone could consume it. 

That imbalance between production and consumption of knowledge is a mental frame worth holding onto.

Bush goes further: he focuses on associative learning, the way memory connects ideas by association. He imagines voice-controlled systems, personalized learning, and interactive knowledge navigation. All of this in 1945, when the first computers were just being built.

Why AI safety breaks at the system level

AI safety shifts from the model to the system level. As AI becomes agentic and tool-driven, risk emerges from complex interactions, widening the gap between evaluation and real-world behavior.

AI Accelerator InstituteAndrew Lovell

From Bush to Engelbart to Jobs

Bush’s vision was picked up by Douglas Engelbart in 1963, who created A Conceptual Framework for the Augmentation of Man’s Intellect and spent his life building toward it. 

In 1968, he delivered what’s now called the “mother of all demos” – over an hour and a half of live demonstration that introduced the personal computer, Windows, hypertext, graphics, the computer mouse, word processing, video conferencing, and collaborative real-time editing. 

A recent PwC study dropped a stat worth jotting down on a Post-it: 74% of AI’s economic value is currently captured by just 20% of organizations.

The remaining 80% are generating activity (dashboards, proofs-of-concept, enthusiastic all-hands updates) while producing disproportionately modest returns. 

If your organization has been in “pilot mode” for 18 months, this article is personally addressed to you…

What the data actually shows

PwC’s 2026 AI Performance Study surveyed 1,217 senior executives across 25 sectors, measuring revenue and efficiency gains attributable to AI against industry medians.

The methodology filters out the most common form of AI reporting: claiming credit for improvements that would have happened anyway. 

What remains is a stark, widening gap between a small cohort of leaders and a majority still perfecting their pilot-to-production PowerPoint transitions.

The behavior gap is bigger than the technology gap

The AI leaders are boundary-crossers competing in adjacent markets, and their returns are outpacing efficiency-focused deployments by a margin that is becoming hard to explain away.

What separates the top performers

The gap is less about model selection or prompt engineering and more about what the AI is actually pointed at. 

Spoiler: it is pointed at revenue, autonomy, and new markets, rather than shaving 12% off the accounts payable process. The behaviors driving performance are structural, replicable, and conspicuously absent from most AI roadmaps.

The practices separating leaders from the rest:

• Autonomous decision-making at scale: Leaders are 2.8 times more likely to have increased decisions made with full automation, backed by governance structures that make that autonomy trustworthy rather than just fast.
• Growth over cost reduction: The leading cohort treats AI as a reinvention engine, directing it at new market entry and revenue expansion rather than internal efficiency theater.
• Governance as a scaling prerequisite: High performers build evaluation and monitoring infrastructure before scaling, moving faster because they invested in foundations first, a sequencing insight that most roadmaps quietly reverse
• Cross-sector collaboration: Leaders combine AI with external partner strengths to unlock use cases that single-sector competitors are structurally unable to replicate

3 easy ways to get the most out of Claude code

Everyone is talking about Claude Code. With millions of weekly downloads and a rapidly expanding feature set, it has quietly become one of the most powerful tools in a developer’s arsenal. But most people are barely scratching the surface.

AI Accelerator InstituteAndrew Lovell

Why the majority are stuck in pilot purgatory

The dominant adoption playbook (start low-risk, build confidence, expand gradually) is producing learnings ahead of returns for most organizations.

Teams cycling through proofs-of-concept often ask “which use cases should we prioritize?” when the binding question is “what would our data infrastructure need to look like for AI to compound?” 

Those are different problems, and the second one requires slightly more than a new Jira board.

That reframe is where the top 20% started, and the majority have yet to arrive.

What practitioners should actually do with this info

The data makes it reasonable to ask whether gradual expansion is delivering what it promised. 

For organizations still “building internal confidence” two or three years in, the answer the data suggests is: probably less than you reported upward. The levers available are structural, and the sooner they are pulled, the wider the compounding gap becomes.

AI agents: building trust via cryptographic proof

As AI agents grow more autonomous, trust can’t rely on logs alone. In this this article, I explore how cryptographic techniques — from content-addressed code to tamper-evident audit trails — are laying the groundwork for a new era of verifiable, auditable AI.

AI Accelerator InstituteAsutosh Mourya

Practical shifts worth prioritizing:

• Reframe the success metric: Measuring AI by cost reduction optimizes for the wrong variable; leaders measure revenue attributable to AI, new markets entered, and decisions automated at acceptable error rates.
• Invest in foundations before scaling pilots: Governance, data quality, and model evaluation pipelines are prerequisites for compounding returns: scheduling them for ‘next quarter’ is how pilot programs generate the illusion of progress.
• Find convergence opportunities deliberately: Cross-sector growth requires explicit effort to identify where AI capabilities combine with external partner strengths to create something each party would struggle to build independently.
• Separate learning investments from return investments: Both are legitimate, but conflating them is how organizations stay permanently impressed by their own pilots while the top 20% widen the gap further.

PwC’s conclusion is direct

The performance gap will keep widening as leaders learn faster, scale proven use cases, and automate decisions at scale. For practitioners, that framing should feel clarifying rather than alarming. The gap is a structural consequence of strategy, and strategy is something organizations can change.

*Source: PwC 2026 AI Performance Study, published April 13, 2026*

AI systems can now execute arbitrary tasks autonomously — running code, invoking external APIs, and making decisions without direct human oversight. 

This creates a foundational trust problem: when an agent acts independently, how do you know the results are accurate, repeatable, and untampered with? For regulated or mission-critical environments, these questions demand concrete answers.

The EU’s proposed AI Act requires traceability and tamper-evident logging for all high-risk AI systems. Yet most agent workflows still rely on standard log entries or short-lived records that can be easily forged or altered by malicious actors or faulty system components. 

As one industry expert observed, AI systems can generate code faster than any team can review it, making new approaches to validating programmatically generated outputs essential.

Trust must be built on a new base-layer approach that:

• Binds data and code together through cryptographic means
• Ensures deterministic processing on every run
• Provides an unalterable history of all actions taken

Content-addressed artifacts

Immutability is central to verifiability. All code and models an agent uses should be linked to a cryptographic hash, treating tools, skills, and prompts as content-addressed artifacts with Content IDs (CIDs).

Any modification creates a new CID, instantly breaking downstream references and making unauthorized changes immediately detectable. 

An agent’s full identity — including model versions, library versions, and skill definitions — can be expressed as a set of hashes or signatures, so any attempt to load a malicious code module fails immediately on hash mismatch.

ContextSubstrate puts this into practice by documenting each agent run as an immutable context package tied to a SHA-256 hash. 

Every input, parameter, interim step, and output is stored in a single content-addressable bundle with a unique context URI (e.g., ctx://sha256:…). 

Deterministic and reproducible inference

Content-addressing fixes what code runs; determinism ensures it produces the same result every time. Modern LLMs have traditionally been non-deterministic, but recent research shows this is not an inherent constraint:

• Karvonen et al. found that using fixed random seeds and sampling parameters produced identical tokens in approximately 98% of cases across repeated runs.
• EigenAI demonstrated true bit-for-bit deterministic inference on GPUs by carefully controlling the execution environment and removing all sources of non-determinism, achieving identical output byte streams on every run.

EigenAI paired this with a blockchain-style cryptographic log — encrypting and recording all requests and responses on an immutable ledger. 

Verification then reduces to a simple hash comparison of the output, giving every model prediction a self-contained proof of correctness.

Where full determinism is not achievable, reproducibility commitments offer a practical alternative. 

An agent declares that its results will be deterministic within an acceptable variance boundary, and a verifier can later confirm this by replaying the run with the same seed, prompt, and model configuration. 

Code generation tasks tend to be fully repeatable; more variable outputs can be assessed using semantic equivalence comparisons or thresholded edit distance.

AI’s new era: Train once, infer forever in production AI

Why the future of AI systems will be driven by inference and agent workloads.

AI Accelerator InstituteAmitChaudhary

Run-time isolation and sandboxing

Reproducibility addresses the integrity of outputs; isolation constrains what an agent can do in the first place. As NVIDIA’s AI Red Team notes, AI coding agents often execute commands with the user’s full system privileges, vastly expanding the attack surface. A compromised or errant agent could:

• Write to critical system files
• Exfiltrate sensitive data
• Spawn unauthorized rogue processes

The practical guidance is to treat all agent tool-calling as untrusted code execution. Key mandatory controls include:

• Blocking all unapproved network egress to prevent unauthorized external connections or data exfiltration
• Confining file-system writes to a designated workspace, disallowing access to sensitive paths such as ~/.zshrc or .gitconfig
• Dropping root privileges and applying kernel-level isolation via secure runtimes like gVisor or Firecracker microVMs, OS sandboxing tools such as SELinux or macOS Seatbelt, or eBPF/seccomp filters

WebAssembly (Wasm) offers a compelling lightweight option: a portable bytecode sandbox with no system calls by design. 

Agent code compiled to Wasm can only access explicitly granted host functions, eliminating the shared-kernel risks of traditional containers. Combined with memory and time limits, Wasm provides a powerful execution environment for generated scripts and tools.

The principle holds: autonomy should be earned through demonstrated trustworthiness, not granted by default.

Why AI safety breaks at the system level

AI safety shifts from the model to the system level. As AI becomes agentic and tool-driven, risk emerges from complex interactions, widening the gap between evaluation and real-world behavior.

AI Accelerator InstituteAndrew Lovell

Tamper-resistant logging and proof bundles

Isolation and determinism control what agents do; logging provides accountability for what they did. Standard logs lack cryptographic linkage, meaning entries can be removed or altered without detection. 

A better solution is an append-only, Merkle-chain audit trail where each log entry’s hash is chained to the previous one — any deletion or modification breaks the chain immediately.

Zhou et al.’s Verifiable Interaction Ledger takes this further: every agent-tool transaction is both hashed and bilaterally signed by two parties, meaning no entry can be secretly added or modified.

Conclusion: toward a trustworthy agent ecosystem

Verifiable execution applies established techniques — content hashing, reproducible builds, and sandbox confinement — to LLM agents, creating a multi-layered trust framework where:

• Agents are tied to specific code sets via digitally signed certificates
• Models run deterministically under fixed random seed conditions
• Every step occurs within a hardened, isolated sandbox
• All interactions are recorded in a tamper-evident hash chain

The result is full auditability: any party can replay the sequence of hashes and verify that an agent’s actions were consistent with the original intent and declared policy.

3 easy ways to get the most out of Claude code

Everyone is talking about Claude Code. With millions of weekly downloads and a rapidly expanding feature set, it has quietly become one of the most powerful tools in a developer’s arsenal. But most people are barely scratching the surface.

AI Accelerator InstituteAndrew Lovell

The momentum behind this approach is real. 

Academic work — including the VET and Genupixel frameworks — has formally characterised chainable verification. Commercial SDKs are beginning to emerge, and regulatory pressure from the EU AI Act is pushing organizations to demonstrate tamper-resistant logs and reproducibility for high-risk AI uses.

The black-box era of agentic AI is coming to an end. It will be replaced by a paradigm where every autonomous decision carries a verifiable proof of integrity — from content-addressed code to digitally signed audit trails. 

As AI agents take on more of our digital work, this verification layer will be the essential safeguard against error, manipulation, and loss of confidence.

What happens when advanced AI capabilities enter the cybersecurity stack at scale?

For AI professionals, this marks an important moment. Cybersecurity AI now sits at the intersection of infrastructure, governance, and real-world application.

In other words, it has moved from interesting to essential.

So what does this mean for AI professionals?

The rise of AI-native cybersecurity tools

AI-driven cybersecurity continues to evolve from passive detection into active analysis and response. Models such as GPT-5.4-Cyber introduce capabilities that extend far beyond traditional tooling.

Security teams now have access to systems that can interpret compiled binaries, identify anomalies, and surface vulnerabilities without requiring source code.

This represents a meaningful acceleration in workflows that previously required manual reverse engineering and deep domain expertise.

The result is a shift toward AI-augmented security operations, where analysts operate alongside models that continuously evaluate and interpret complex systems. The coffee consumption may stay the same, yet the output per analyst looks very different…

Why AI safety breaks at the system level

AI safety shifts from the model to the system level. As AI becomes agentic and tool-driven, risk emerges from complex interactions, widening the gap between evaluation and real-world behavior.

AI Accelerator InstituteAndrew Lovell

Two emerging approaches to access

As these capabilities mature, different deployment strategies are taking shape. The contrast reflects a broader design decision within AI cybersecurity.

Some platforms emphasize controlled distribution, where access is limited to a small group of verified organizations. This approach prioritizes tight oversight and curated usage environments.

Others adopt a broader access model, where entry is granted through identity verification and structured onboarding. This approach focuses on enabling a wider pool of security professionals to leverage advanced tools.

What this means for AI professionals

For practitioners, access models now play a central role in how cybersecurity systems are integrated into existing workflows. The conversation has expanded from capability evaluation into deployment strategy.

Security leaders and AI engineers increasingly evaluate questions such as:

• How AI tools integrate into existing security pipelines and SIEM platforms• How identity verification frameworks support controlled access at scale

• How model outputs align with internal validation and audit processes

• How teams manage collaboration between human analysts and AI systems

These considerations highlight a broader trend. AI cybersecurity requires alignment across engineering, security, and governance functions. Silos rarely perform well under pressure, and as we all know, cybersecurity provides plenty of pressure.

3 easy ways to get the most out of Claude code

Everyone is talking about Claude Code. With millions of weekly downloads and a rapidly expanding feature set, it has quietly become one of the most powerful tools in a developer’s arsenal. But most people are barely scratching the surface.

AI Accelerator InstituteAndrew Lovell

The operational impact on security teams

AI-powered cybersecurity tools introduce measurable improvements in speed and coverage. At the same time, they reshape how teams approach daily operations.

Routine analysis tasks can be automated or augmented, allowing analysts to focus on higher-value investigations. Pattern recognition and anomaly detection benefit from continuous model evaluation, providing earlier visibility into potential threats.

At the same time, teams gain the ability to inspect complex systems with greater depth. Reverse engineering, malware classification, and vulnerability detection become more accessible across a wider range of skill levels.

This evolution supports a more distributed model of expertise, where advanced capabilities extend across the organization rather than remaining concentrated in specialized roles. More eyes on the problem, fewer bottlenecks in the process.

Key considerations for implementation

As organizations adopt AI-driven cybersecurity tools, several practical considerations come into focus:

• Integration: Alignment with existing infrastructure, including cloud environments and security platforms

• Validation: Processes for verifying model outputs and ensuring reliability in high-stakes scenarios

• Access control: Mechanisms for managing user permissions and maintaining secure usage

• Monitoring: Continuous oversight of model behavior and system performance

These factors shape how effectively AI systems contribute to security outcomes. Strong implementation frameworks support both performance and trust.

Building trust in AI-driven security systems

Trust remains a central component of AI adoption in cybersecurity. Teams rely on systems that operate consistently, transparently, and with measurable accuracy.

Clear audit trails, reproducible outputs, and well-defined evaluation metrics contribute to confidence in AI-generated insights. Structured access models further support trust by ensuring that usage aligns with organizational policies and standards.

As AI systems take on more responsibility within security workflows, trust becomes an operational requirement rather than a conceptual goal.

AI’s new era: Train once, infer forever in production AI

Why the future of AI systems will be driven by inference and agent workloads.

AI Accelerator InstituteAndrew Lovell

Looking ahead: Access as a design decision

AI cybersecurity continues to evolve rapidly, with new models and capabilities entering the landscape at a steady pace. Alongside this growth, access models have emerged as a defining factor in how these systems are used.

For AI professionals, this represents a shift in focus. Technical capability remains essential, while deployment strategy now carries equal weight. Decisions around access, verification, and integration shape how effectively AI contributes to security outcomes.

The next phase of AI cybersecurity development will likely bring further innovation in both capability and delivery. Teams that approach access as a core design decision will be well-positioned to adapt and scale.

Innovation in AI cybersecurity continues to accelerate. With the right access models in place, organizations can translate advanced capabilities into practical, high-impact security outcomes.

And ideally, sleep a little better at night…