The agent adoption ladder: personal → team → enterprise, without the faceplant

Every organization’s agent story starts the same way: someone automates a piece of their own drudgery — often with something like the triage agent — it works, a teammate wants it, and suddenly a personal script is infrastructure. Each rung of that ladder changes the engineering contract. Climbing deliberately is cheap; discovering the changes in production is not.

Rung 1 — Personal: the only rule is honesty with yourself

One user, one machine, blast radius of one. Keep it loose, with three habits that make the next rung cheap:

• A digest — the agent reports what it did; you never wonder.
• Reversibility — it drafts, moves, and labels; it doesn’t send, delete, or pay.
• A data line you don’t cross — no customer data, no credentials in prompts, even on your own machine. (This is the habit that saves you later.)

Rung 2 — Team: the jump everyone underestimates

The moment a second person depends on it, four things change at once:

1. Identity. “It runs as me” breaks — the agent needs its own service identity and scoped credentials, or one teammate’s departure kills the team’s tooling.
2. Secrets. Out of scripts and .env files on laptops, into a secret manager, held by the tool layer — never the prompt.
3. Review. Prompts, tool definitions, and hooks go into version control with pull-request review. The agent’s configuration is code now, and which mechanism does what tells you who reviews what.
4. An owner and an undo. One named person answers “it did something weird,” and every write action has a documented reversal path.

The team rung’s rule of thumb: a personal agent needs a digest; a team agent needs a digest, an owner, and an undo.

Rung 3 — Enterprise: multiply by every team

At company scale the problems are portfolio problems: forty teams independently discovering the same four lessons, forty gateway bills, forty security reviews. The enterprise moves are consolidation moves:

• One gateway for model access — keys, budgets, and cost attribution in one place.
• A paved road: approved models, a starter template, an eval harness, and the tool risk matrix — so the easy path is the safe path.
• Tiered oversight by action class via approval patterns, instead of one blanket policy that’s too strict for readers and too loose for writers.
• An inventory: who runs which agent, with which tools, touching which data. For most industries this is good hygiene that pays off in incident response and audits alike; the governance framework is the full treatment when you need it.

The anti-patterns at every rung

• Skipping rungs — a personal script promoted straight to customer-facing because a demo went well.
• Blocking rung 1 — banning personal agents doesn’t stop them; it just removes your visibility. Publish the paved road instead.
• Enterprise ceremony at team scale — a three-person team doesn’t need a review board; it needs an owner, a repo, and an undo.

The ladder’s entire point: match the ceremony to the blast radius, and make the next rung cheap before you need it.