Solution · AI Operations

Every AI agent you run has a switch-on date. Almost none has a review date or a switch-off date.

Building an agent is a project. What almost nobody has built is the function that governs it afterwards: a catalog where every agent has an owner and a version, a gate that decides which change reaches production, a rollback that works, a review that checks whether it still earns its keep, and a real retirement — credentials revoked. We build that lifecycle and we run it.

Ship it in my business See how it works

The problem

Everyone remembers agent number one. Nobody knows who asked for number fourteen, what version it runs, or whether anyone uses it.

Nobody has the list. There is no single place holding every agent running today with its owner, its version, its permissions and the reason it was approved: there is a stale spreadsheet and the memory of three people.
A prompt change reaches production because someone edited it, not because it passed a gate: no test environment, no evaluation comparing it against the previous version, and no clean way back if it turns out worse.
“Version” means different things depending on who you ask: the model changed on its own, the prompt was hand-edited, and the connected tools are not the ones from three months ago — and none of it is pinned to a number you can name.
There are agents nobody uses anymore, still switched on: burning tokens every day and keeping write-scoped credentials alive, and nobody shuts them off because nobody is sure what breaks if they do.
When someone asks whether an agent still earns its keep, the answer is an opinion: no baseline, no review cadence, and no written criterion for when it gets rebuilt and when it gets retired.

Cost of staying the same

Gartner predicted, in its press release of 25 June 2025, that over 40% of agentic AI projects will be canceled before the end of 2027 — and the reasons it gives are not technical: escalating costs, unclear business value and inadequate risk controls. All three are managed in the lifecycle or they are not managed anywhere. And the gap has been measured: in the Cloud Security Alliance whitepaper “The Non-Human Identity Governance Vacuum” (May 2026), 78% of organizations report having no documented policy for creating and deleting AI identities, 51% have no clear owner for those identities, and only 20% have a formal offboarding process with key revocation. Translated to your company: every agent you switch on without a lifecycle is cost nobody revisits and a credential nobody will revoke. It does not fail on day one. It accumulates.

The solution

We build your agent lifecycle — catalog, promotion gate, versioning with rollback, review and retirement — and we run it as a continuous function

1We build the real catalog, not the one you think you have: which agents run today, who asked for them, who answers for each one, which version is live, on what permissions and with which credential. Agents nobody remembered switching on show up almost every time.
2We put in states and gates. An agent goes from proposed to validated, from validated to production and from production to retired along a written path, not because someone deployed it on a Thursday. Before production there is an evaluation against real cases from your business, and it is that evaluation that opens the gate — not the feeling that it looks fine.
3We make “version” mean something: prompt, model, connected tools and limits get pinned to a named version, comparable against the previous one and revertible in minutes. It is what the serious platforms already do — AWS Bedrock AgentCore freezes each update as an immutable version and rolls back by repointing the endpoint; Microsoft Foundry keeps agent snapshots and promotes across environments through evaluation-driven gates — and we build it on what you already run instead of bringing you another tool.
4We add the periodic review, which is where you decide whether an agent stays alive: what it costs, how much it is used, how it performs against its baseline and whether the process it served still exists. With a written criterion for when it gets tuned, when it gets rebuilt and when it gets retired — so the decision does not depend on who is in the room.
5We close at the end, which is the piece almost nobody has built: retirement. The agent is switched off, its credentials and access are revoked, you decide what is kept from its logs and for how long, and whatever it did is handed over. It is the retirement stage an AI management system like ISO/IEC 42001 requires you to handle explicitly — and, incidentally, the one that stops you paying for agents nobody uses.

What changes

What you stop losing

The catalog stops being a stale spreadsheet: every live agent has an owner, a version, permissions and the reason it was approved — and what is not in the catalog does not run.
Mechanism
A bad change stops being an afternoon of work: the previous version has a name and you go back to it, instead of rebuilding by hand the prompt you had on Tuesday.
Mechanism
The agent nobody uses stops being invisible spend and an open credential: the review catches it and retirement actually closes it, with access revoked.
Mechanism
What we measure: % of live agents with owner and version in the catalog, % of changes that reached production through the evaluation gate, rollback time to the previous version, agents reviewed and retired per quarter, and credentials revoked after each offboarding.
What we measure

Spec sheet

Work it removes: finding out too late which agents run, who asked for them and which version is live — and being unable to roll back a bad change or fully switch off the one nobody uses
Typical setup: 4–8 weeks
Input: the agents you already have running, with their prompts, their models, their connected tools and their credentials
Output: a live catalog with owner, version and state per agent, a promotion gate with evaluation, rollback to the previous version, periodic review against written criteria, and a retirement that revokes access
Works with: AWS Bedrock AgentCoreMicrosoft Foundry Agent ServiceGoogle Vertex AI Agent EngineLangSmithLangGraphn8nOpenAIAnthropic
Can connect to: Your version control and your CIYour secrets manager and your identity providerYour evaluation suite over real business casesYour non-human identity inventory, if you already keep one
What we measure: % of live agents with owner and version in the catalog% of changes promoted through the evaluation gaterollback time to the previous versionagents reviewed and retired per quartercredentials revoked after each offboarding
Good fit for: companies that have gone from their first agent to their tenth and are starting to lose track of how many they have, who answers for each one, and which could be switched off without anyone noticing
Not a fit for: anyone running a single agent in production with one person who knows it end to end: there, the lifecycle is a conversation, not a function

Frequently asked questions

In the object. Maintenance answers “keep my agents working well”: watch for drift, update when a model changes, control cost, react when something degrades. You have that covered in who maintains AI agents in production, and responding to a specific failure is incident management. The lifecycle answers something else: which states an agent passes through and who authorizes each jump. Onboarding with an owner and a reason, validation before production, promotion through an evaluation gate, named versions with rollback, review on a cadence, and retirement with revocation. Maintenance is the verb of the day-to-day; the lifecycle is the frame that decides what comes in, what changes and what goes out. Without it, maintenance ends up nursing things nobody should still have switched on.

It gives you the mechanism, not the function, and that difference is what costs money. AWS Bedrock AgentCore versions immutably and lets you revert; Microsoft Foundry keeps snapshots and lets you put evaluation gates between environments. Good: we use them. What no platform gives you is who answers for each agent, which evaluation a change has to pass to be acceptable in YOUR business, how often you check whether the agent still earns its keep, on what criteria you decide to retire it, or what happens to its credentials and its logs the day it is switched off. Those are decisions and cadences, and someone has to run them. And if your agents are spread across several platforms — normal past the third one — the catalog and the criteria have to live above all of them.

Switching it off is the easy part and the one that matters least. Real retirement is revoking its credentials and permissions — which is where almost everyone fails: only 20% of organizations say they have a formal offboarding process with key revocation, per the Cloud Security Alliance whitepaper of May 2026 — deciding which logs are kept and for how long, telling whoever depended on it, and handing over what it did. A “switched-off” agent whose identity is still live is not a retired agent: it is a write-scoped credential nobody watches anymore because, officially, that agent does not exist. That is why retirement is a lifecycle stage with a checklist, not a switch.

Want it running in your business?

You’ve pinned the problem. We ship the fix and leave it measured.

See the service