N100FN100F
Published on

Agents that execute, people that decide

AI assistants already write code, deploy and document. Inside a governed platform they can do a great deal of work autonomously, as long as the gates that require judgement stay with people.

Service line Platform Engineering

A new kind of team member

An AI assistant with access to a repository and a terminal can today build an application, deploy it to a test environment, write its documentation and fix its own errors. It does in hours what used to take days.

The temptation is to let it do everything. The risk is the same as with any new person given too many permissions: that it does something irreversible with a lot of confidence and not enough context.

Dividing by type of decision

Agents that execute, people that decideThe agent lane that executes and the person lane that decides at the gates.PLATFORM · AIAgents that execute, people that decideAGENT · EXECUTESPERSON · DECIDES Builds Tests Deploysto acceptance Documents Updatesstatus Approvesthe specification Validatesto acceptance Authorisesproduction Irreversiblechangesthe agent stops and waitsThe agent’s permissions match the division:it has no credentials for what it must not do alone.N100F · n100f.com

The way of working that does work divides the labour by type of decision, not by difficulty:

The agent executes, autonomously:

  • Building and testing according to the approved specification.
  • Deploying to development and acceptance environments.
  • Writing and updating the documentation, the status and the logbook.
  • Fixing whatever the automatic validations reject.

The person decides, always:

  • Approving the specification (what is going to be built).
  • Validating in acceptance (is this what was asked for?).
  • Authorising the move to production.
  • Any irreversible change: deleting data, touching the public network, changing domain records.

What makes it work

  • The gates are written down and the agent knows them: it knows where it has to stop and wait.
  • The permissions match the division: the agent has no credentials for what it must not do alone.
  • Everything is recorded: what the agent did appears in the logbook just like what a person does.
  • The agent explains before acting on anything that affects others: what it is going to change and why.

The division works if the agent knows precisely what is right and what is not. That is not the agent's call: it is decided by the rules people write.

How we apply it: Platform Engineering.

If your cloud cost keeps growing and nobody knows why, or your architecture depends on whoever built it, let’s talk.

Let’s talk