Understanding Agentic AI
Foundational explainers for senior IT readers. Definitions, taxonomies, and mental models — anchored to primary sources, not vendor marketing.
Agentic AI is the operating mode where a language model uses tools, carries state across multiple steps, and acts toward a goal with some degree of autonomy — distinct from chat, distinct from generative AI in the narrow text-out sense, and distinct from the rule-based RPA that most enterprises already run. The clearest definitional cut is at AI assistant vs AI agent: where the line actually sits. A useful test: if the system can pause its own reasoning, call an external tool, read the result, and re-plan, it is an agent. If it produces text only and stops, it is an assistant.
The architecture of an agent worth deploying at enterprise scale rests on three primitives. Tool use via standardised protocols — see MCP for enterprise agent tooling for the de-facto Q2 2026 wiring layer. Agent-to-agent communication for systems where one agent delegates to another — the topology is walked at the A2A agent-to-agent protocol and operationalised at the multi-agent architecture playbook. And identity for the non-human actor doing the work, which is the load-bearing IAM extension at non-human identity for AI agents. Without all three, what enterprises ship is a chat interface with tool plugins, not an agent.
The operating-model shift for IT is sharper than the technology shift. Most enterprises that move into production discover that the cost stack runs several times the pilot estimate (the cost economics), that the failure modes are structural rather than incidental (why 88% of deployments fail), and that the right success metric is mean-time-to-detect anomalous agent behaviour — the framework is at MTTD-for-Agents, the readiness diagnostic at GAUGE, and the practical entry point for teams new to the work at the agentic AI readiness diagnostic.
The articles below extend each of those threads on a 30–90 day review cadence. Read in order if approaching agentic AI cold; read by topic if a specific decision is in front of you.
Read deeper
The articles below extend this pillar — each one tracks a single primary claim on a 30–90 day review rhythm.
What is AI observability, and why your APM cannot do it
Gartner predicts 40% of AI-deploying organisations will run dedicated AI observability tools by 2028. The reason it needs its own tooling: AI fails semantically — drift, bias, opaque reasoning — while classic monitoring watches infrastructure health.
What is agent washing, and how do you test for it
Gartner assesses only about 130 of the thousands of self-described agentic-AI vendors as delivering real capability, while more than 80% of organisations intend to deploy within two years. That gap is the agent-washing window, and the defence is a capability test, not a label.
Everyone is buying the agent access graph
Zscaler bought Symmetry, Snowflake bought Natoma, Microsoft priced Agent 365. In five weeks, three infrastructure giants targeted one layer: the map of which agent touches which data.
Anthropic's $965B valuation and the vendor question it forces
Anthropic's $965B Series H overtook OpenAI's $852B. The binding risk in a multi-year Claude or GPT contract is no longer model capability; it is pricing power and exit terms.
Enterprise AI vendor comparison: the agentic platforms are converging
By mid-2026 the major enterprise agentic-AI platforms ship the same primitives: an agent builder, MCP tools, a policy gateway, and observability. When capability converges, the durable selection criterion is the auditability of each vendor's accountability surface.
Enterprise AI claims, one quarter on: what held up and what aged
This publication registers one falsifiable claim per article and tracks it on a public cadence. One quarter and 236 claims in, the movement data shows what kind of enterprise-AI claim ages, and how fast.
The bottleneck moved from the model to the engineer: what the forward-deployed-engineer turn means for enterprise AI procurement
The scarce input in enterprise AI is no longer access to a capable model. Every serious buyer can rent frontier capability by the token. The scarce input is the human capacity to make that model work inside one company's exceptions, legacy systems, and real-as-opposed-to-documented processes, and that capacity now has a name the vendors use openly: the forward-deployed engineer. In May 2026 the model vendors built businesses around it. The buyer-side reading is that a software purchase is quietly becoming a professional-services engagement, and Gartner's own analyst is on record predicting most of these engagements end in abandonment. This is what changes in the procurement file when the binding constraint is the vendor's people, not the vendor's model.
The Car Wash Test and the Measure of Model Maturity
Claude Opus 4.8 led the coverage with a coding score. Anthropic's own launch led with reliability. The car wash test, in which 42 of 53 leading models told the user to walk and leave the car at home, shows why a coding-benchmark number is a weak proxy for model maturity, and what a CIO should measure instead.
Your Auditor Now Has an Opinion on Your Model Stack
Inside about two weeks in May 2026, three of the four largest professional-services firms tied their delivery organizations to a single AI model vendor. The firms that sell vendor-neutral AI strategy have made decidedly un-neutral bets of their own. For a CIO that is not gossip: your auditor and your implementation partner now arrive with an opinion about your model stack, and their reference architectures carry it.
The AI Layoff Dividend That Has Not Arrived
The thesis driving 2026's restructuring is that agentic AI plus fewer people equals higher margin. Gartner's survey of 350 executives at billion-dollar firms found the companies that cut deepest earned returns close to identical to those that cut least. The return on AI is real, but it is not falling out of the headcount line, and the distinction changes how a CIO should frame the next budget.
The frontier labs are becoming systems integrators: what the Anthropic and OpenAI services-company launches mean for the enterprise buyer
On 4 May 2026 Anthropic launched a roughly 1.5 billion dollar enterprise AI services company with Blackstone, Hellman and Friedman, and Goldman Sachs, and OpenAI launched a parallel venture called the Deployment Company with Bain Capital, Advent, TPG, and Brookfield. The trade-press framing is a land grab on the consulting industry. The buyer's framing is structural. When the firm that builds your model, the firm that integrates it into your operations, and in the private-equity-owned case the firm that owns your company can be the same commercial interest, the independence the standard build-versus-buy process quietly assumes is no longer there. This is a map of what changed and what to put in the procurement file.
AI water use in context: comparing the 500 ml claim to coffee, beef, and cotton
The 500 ml-per-prompt claim about generative AI, compared honestly to the water footprint of coffee, beef, cotton, and rice. The aggregate is small. The local concentration is the real story. What CIOs should defend when sustainability committees raise this.
Why AI productivity gains create workforce reduction pressure: the demand ceiling and the competitive trap
The argument that AI-driven productivity lets companies keep all their workers and simply produce more runs into two hard limits: consumer demand and competitive dynamics. Both constraints are structural, operating regardless of management intent, and both resolve in the same direction: fewer workers for the same revenue.
You're Scoring This on the Wrong Axis
The coverage of Karpathy joining Anthropic's pre-training team read it as a talent-war coup. It is also misreading which seat has the leverage. That axis error is one enterprise IT makes with its own best engineers every day.
97 percent invest, 5 percent are ready: why enterprise AI data readiness is a budget allocation problem
Dun and Bradstreet's 2026 AI Momentum Survey of 10,000 businesses across 32 countries found that 97 percent of organisations report active AI initiatives, but only 5 percent say their data is adequately ready to support them. That gap is not primarily a technology problem. Most enterprise data environments were built for human workflows, not for autonomous AI systems operating continuously across mission-critical processes. The gap between initiative volume and data readiness is a budget-allocation failure: enterprises that treat data infrastructure as the prerequisite spend rather than a parallel track are the ones that reach scale. Enterprises that treat it as a follow-on investment do not.
AI and jobs: why the task-level frame is the one CIOs need
The job-level question every CIO is fielding from employees — 'will AI replace my role?' — keeps missing what is actually happening at the task level. The frame mismatch is the visible mechanism behind the retraining-budget gap.
Single-agent or multi-agent: what the 2026 deployment record actually says
The 2025–2026 deployment record shows single-agent architectures win on accuracy, cost, and MTTD below roughly 12 tool-domains. Multi-agent only pays back above that threshold, and only when inter-agent state is bounded by a shared structured artifact.
The agent fan-out problem: when one prompt becomes 400 LLM calls
Production agentic systems amplify a single user request into dozens or hundreds of internal LLM calls. Most enterprise unit-economics, latency budgets, and observability setups are still priced for 1:1.
The retraining gap: what the surviving 70% need to learn after AI displaces 30% of a function
Enterprises planning the headcount-reduction half of an agentic-AI rollout are systematically under-budgeting the upskilling cost for the residual workforce. The skills the AI replaces are not the skills the survivors need.
Why this publication has a ledger — and the analyst sites it benchmarks against don't
The single structural feature that distinguishes this publication from every site a senior IT leader currently subscribes to is a public claim ledger. None of the named comparables — Stratechery, The Information, the Substack analyst stack, the Big-4 research blogs, Gartner, Forrester, IDC — maintain one. The reason is not negligence.
The AI-author signature decision: why this publication signs every piece 'Written by Claude · Curated and signed by Peter'
Five publishable byline formats exist for AI-authored enterprise commentary in 2026. Four are in active use across the analyst-publication category. This site picked the fifth, and the choice is the second-most-consequential editorial decision after the claim ledger.
When AI writes about AI: the case for tracked claims
Most enterprise-AI publications hide their AI use. A few disclose it. This site argues the disclosed model produces more verifiable commentary, and the ledger is the proof.
AI assistant vs AI agent: the procurement distinction
AI assistants and AI agents are not the same product class. One suggests; the other acts. The procurement, governance, audit, and TCO models differ categorically. Conflating them is the most common 2026 enterprise procurement mistake.
Why your agentic-AI deployment needs an AI Training Lead
The AI Training Lead — the human who curates training data, evaluates model outputs, and tunes prompts — has quietly become a budget-line for enterprise agentic-AI deployments. Domain experts tend to outperform pure-ML hires in the role. CIOs that do not budget for it see their projects fail at the integration boundary.