Skip to content

Governance · · 8 min read

Who Owns the Agent? November’s Releases Are About the Control Plane, Not the Model

On 9 November Anthropic released Claude Fable 5.2 on every major cloud, with a more capable sibling, Mythos 5.2, held back for vetted organisations. On 10 November IBM and Microsoft announced Imagine Enterprise, a platform for building, running and governing AI agents. On 17 November, the opening day of Ignite, Microsoft described Agent 365 as a control plane for agents and said it already holds more than 50 million of them; the same day OpenAI’s GPT-6.2 Astra was reported at exactly the list price of Anthropic’s new model, with a cheaper sibling, Sol, beside it. Read together, the month says something plain: the models have converged, and the work that now separates a useful agent from a risky one is knowing which agents you have, what each may touch, and who is accountable for it.

The inventory problem arrives

September’s story was always-on agents; October’s was the approval gate. November’s is what happens when there are a great many of them. Microsoft’s figure — more than 50 million agents already registered in Agent 365 — is a vendor’s own count, but the direction is not in doubt. Once anyone in a company can start an agent, the company has the same problem it once had with spreadsheets and shared passwords: things doing real work that nobody listed.

50 million+agents Microsoft says are already in Agent 365[1]
6,000+agents and assets IBM Consulting brings to Imagine Enterprise[2]
40organisations working with Microsoft on agent standards[1]

Microsoft describes Agent 365 as “a control plane that extends your existing governance, identity, security and management systems” — in other words, agents are to be managed the way staff accounts and devices already are. IBM and Microsoft put it in the first sentence of their announcement: Imagine Enterprise exists to help enterprises “build, run, and govern” agents, with controls, observability and compliance support built in.

Governance is the product now

It is worth noticing what these companies chose to lead with. Not a benchmark, not a demo of an agent booking travel — a register, a set of controls and a way to see what is running. Imagine Enterprise launches first for financial services, insurance and retail, the industries where a regulator will ask the question before a customer does.

Microsoft also launched a transparency initiative alongside NIST’s work on agent standards, with forty organisations already taking part. Standards work is slow and rarely makes headlines, but it is the usual sign that a technology has stopped being an experiment: people are agreeing on how to describe an agent, what it must disclose and how its actions are recorded.

Capability is being rationed, in the open

The model releases tell the same story from the other side. Anthropic shipped Fable 5.2 to everyone with classifiers that block high-risk cyber activity, and kept Mythos 5.2 for organisations that pass a verification programme in security or the life sciences. OpenAI’s GPT-6.2 was reported with tightened safeguards on higher-risk cyber requests — safeguards the company itself is quoted as saying also restrict some legitimate vulnerability research.

November’s flagship releases, list price per million tokens
ModelReleasedInputOutputAccess
Claude Fable 5.29 NovemberUS$8US$40General, with cyber and biology safeguards
Claude Mythos 5.29 November——Vetted organisations only
GPT-6.2 Astra (as reported)17 NovemberUS$8US$40General, tightened cyber safeguards
GPT-6.2 Sol (as reported)17 NovemberUS$2US$10General

Two things follow for a business. First, the frontier is no longer one model that everyone can buy: what you can use depends on who you are and what you are doing with it, and that is now a published policy rather than a quiet refusal. Second, two rival flagships at an identical price is a strong hint that the model itself is becoming a commodity input. The advantage has moved to what surrounds it.

What a small business should do about it

None of this requires an enterprise platform. A company with five agents needs the same four things as one with five thousand, at a smaller scale.

  • A register: every agent and automation written down — what it does, which systems it can reach, and when it was last reviewed.
  • An owner: one named person answerable for each agent, who is told when it fails and who decides when it is retired.
  • Its own identity: an agent signs in as itself with the least access it needs, never with a staff member’s login or a shared administrator account.
  • A record: each action logged with what the agent was asked, what it did and who approved it, kept long enough to answer a question months later.

The test is simple. If the person who set up an agent left the company tomorrow, would anyone know it was still running, what it could reach and how to switch it off? If the answer is no, that agent is a risk regardless of how well it performs.

How we are building for this

In the workflow systems we build, an agent is a first-class user: it has its own account, its own role and its own entry in the audit log, and the administrator’s screen lists agents beside people — what each is permitted to do, what it did this week, and a switch to suspend it. Combined with the three classes we described in September and October — autonomous, approval-required and forbidden — that gives a business the two answers it will increasingly be asked for: what may this agent do, and who is responsible for it.

This is less glamorous than a new model, and it is where the value now sits. Models will keep getting cheaper and more alike. A clear, inspectable account of who is doing what inside your systems — human or machine — is something only you can build, and it is the part that lets you say yes to the next capable agent instead of wondering whether you dare.

Key takeaways

  • November’s headline was not a model: Microsoft’s Agent 365 and IBM and Microsoft’s Imagine Enterprise both lead with governing agents — inventory, controls and observability.
  • Flagship models are converging on price while access is tiered by who you are; the most capable versions go to vetted organisations first.
  • Whatever your size: keep a register of agents, give each a named owner and its own least-privilege identity, and log every action.

Sources

  1. Ignite 2026: Building the frontier ecosystem — Microsoft, 17 November 2026
  2. IBM and Microsoft announce a new AI agent platform designed for industry-specific transformation — IBM Newsroom, 10 November 2026
  3. Claude Fable 5.2 and Claude Mythos 5.2 — Anthropic, 9 November 2026
  4. GPT-6.2 Astra and Sol launch — LLM Stats (secondary; reporting OpenAI’s announcement), November 2026

Choose how to get started

Build with us

A system engineered around how you work.

Scoped in a proposal, designed as a prototype, delivered in milestones you can review.

Start a project

See the work

Systems already running in production.

Marketplaces, operations portals and mobile apps, with the case studies behind them.

View our work
WhatsApp