How it works

Where an agent stops

This figure shows where work leaves my company and where I sign it out.

How the agents ended up needing this map: It took my AI team 17 hours to reinvent middle management.

What you see

Work comes in from outside and goes back out. In the Human lane I set the intent: goals, priorities, limits and what counts as wrong. Intent flows down to the outcome agents, each of which owns one outcome, with task agents beneath them. What they produce rises through automatic code checks, which are not a person, to sign-off. I sign messages, commitments and releases before they leave. I also sign changes to systems and access: the agents propose, I sign, and only then does the change land. What I see at sign-off feeds back into the intent. Every agent reads and writes the knowledge lane, and every agent is built from the foundation of skills, models and systems.

What each control means

• Sign-off. Nothing leaves the company, and no change to systems lands, without my signature.
• Code checks. Automatic checks on the work before it reaches me. They catch what a rule can catch. They do not replace the signature.
• Input review. What a client sends in is checked for privacy and quality before the agents use it or it reaches the records.
• Harness. Code between the agents and the systems they work in. It reaches only the routes on a list, records every write and can undo it. What the agents must never do has no route at all, and a test fails if one appears. The system stays the record: where the log and the system disagree, the system wins.
• Records. Agents add and enrich; they do not delete or merge records without me. There is no guard yet against posting to a closed period: I check dates by hand until the year-end close.
• Memory. Agents write to it without review, so I mostly read it through them.

Where it is still open

Data the agents pull themselves, from the web and APIs, comes in without review. That is how it works today, so the figure shows it. Whether to allow only a list of sources is a separate decision, and I have not made it. Memory is unreviewed too. This is a design and a set of rules I have set, not an audit of how every agent behaves.

Accountability is not in the picture, because it does not move: a person answers for the outcome. Who does the work, who may decide, who answers for it? separates the three, and Why this boundary? explains where I draw the line.

Zoom in

• The Agents lane: What makes an agent an agent · Published
• The Knowledge lane: The firm's memory, shared and individual · Planned
• The split between outcome and task agents: Orchestration: one agent per outcome · Planned