How it happens
Few organisations plan to have many AI agents. It happens gradually. One team uses Claude, another prefers ChatGPT, Microsoft 365 users have Copilot, developers use coding agents, and someone builds an automation that runs overnight. Each was a sensible choice. Together they create a new management problem.
The problems that appear
- No inventory. Nobody can list every AI tool connected to company systems.
- Inconsistent rules. One tool can send email; another can only draft. Nobody decided that; it depends on how each was set up.
- Credential sprawl. Each agent holds its own keys to the same systems.
- Duplicated work. Two agents, used by two people, do the same analysis because neither knows about the other.
- Lost context. Decisions made through one agent are invisible to the rest.
- Scattered records. Each tool logs differently, if at all.
Principles for managing many agents
1. Keep an inventory
For each agent: what it is, who uses it, what it connects to, what it can do, and who is responsible for it.
2. Give each agent an identity
Agents should be distinguishable from each other and from the people they act for. That makes rules, audit and revocation possible per agent.
3. Apply the same rules everywhere
A rule such as “AI may not send external email without approval” should apply to every agent, not only the ones configured carefully. That is easiest when rules are enforced in one place rather than in each tool.
4. Share state, not copies
Agents working on the same client or project should read from and write to the same record of decisions and status, within their permissions.
5. Centralise credentials
The fewer places hold keys to your business systems, the fewer places can leak them.
6. One audit trail
You should be able to ask “what did any AI do on this client last week?” and get one answer.
7. Assign ownership
Every agent should have a named person accountable for it.
Agents that talk to agents
Agents increasingly hand work to other agents. When that happens, each step should narrow what is allowed, and the final action should still be traceable to the person who started it. Be wary of set-ups where one agent can grant another more access than it has.
A starting checklist
- List every AI tool and agent connected to company systems.
- Identify which hold their own credentials.
- Write down the rules that should apply to all of them.
- Choose where those rules will be enforced.
- Decide where shared state and audit will live.
How Kroy approaches it
Kroy gives every agent the same front door. Claude, ChatGPT, Copilot, Codex and your own agents connect to Kroy, each identified separately and linked to the person it acts for. The same policies apply to all of them, they share the same organisational state within their permissions, and every request is recorded in one place.