BvLogic
Diagram · AI PMO

AI PMO Routing Diagram

How a request moves through an AI organisation — entry point, decomposition, forced routing through project-manager roles, and the two structural gates that stop work reaching an unprepared agent.

Download the diagram Read the full guide

SVG. No sign-up, no email.

Most diagrams of multi-agent systems show a manager box with arrows to worker boxes. That hides the part that decides whether the thing works: what is enforced structurally versus what is merely instructed. An agent told to route work through a project manager will mostly do it. An agent that has no path to anywhere else always does.

The shaded steps below are the ones enforced in code. Everything else is a convention, and conventions in agent systems hold about as well as they do in human ones.

How a request reaches an agent that can act on it Entry: Human request (the owner, never an agent) → Entry agent (single named point of arrival) → Decompose (one request to N tasks). Route: Match task to role (against the org definition) → Gate: project work? (if yes, must pass a PM,role) → Gate: role trained? (untrained cannot receive,work) → Assign (task queued to one role). Do the work: Claim (role context assembled) → Shared corpus injected (one source, every role) → Produce deliverable () → File and report up (written to disk, chain,notified). Entry Human request the owner, never an agent Entry agent single named point of arrival Decompose one request to N tasks Route Match task to role against the org definition Gate: project work? if yes, must pass a PM role Gate: role trained? untrained cannot receive work Assign task queued to one role Do the work Claim role context assembled Shared corpus injected one source, every role Produce deliverable File and report up written to disk, chain notified one task at a time Enforced in code, not by instruction Generated from one structural definition
Shaded steps are enforced structurally — the path does not exist otherwise. The two gates are where earlier versions failed: work went direct, and untrained roles answered anyway.

How to read it#

The entry point is single and named. One place a request arrives, which decides distribution. Without it, every requester picks whichever role sounds relevant and the organisation's structure becomes decorative.

The two gates are the load-bearing parts. Both exist because the earlier version failed without them:

  • Project work must pass a project-manager role. Instructed to do this, agents sometimes went direct. Making it structural — there is no other path — ended it. Same principle as spend caps in AI agents: a constraint an agent can reason past is not a constraint.
  • An untrained role cannot receive work. An agent with no role knowledge still answers. It answers generically, and the output is indistinguishable in shape from a good one — which is worse than a refusal.

Shared context is injected, not copied. Every role inherits the same organisational knowledge from one source. Copy it into each role and you have as many places to fix a fact as you have roles, and they diverge immediately.

The organisation is generated, not maintained. The whole structure comes from a single definition, and the build refuses to write on a duplicate identity or an orphaned unit. A hand-maintained agent hierarchy is stale within a month, and stale structure routes work to roles that no longer exist.

What the diagram does not show#

The quality of what comes out. Routing correctly to a well-defined role does not make the deliverable right — and the system produces confident, well-structured work whether or not it had the facts. That limitation, and the uneven depth of role knowledge behind it, is covered on AI PMO and AI Research.

The diagram is an accurate picture of the plumbing. It is not evidence the water is clean.

Back to AI PMO