Home · Agents

How it works: the Project Manager

For keeping a studio of AI agents working as one team: the same rules, shared tools, and every decision the owner needs to make in one place.

Philosophy

  • A rule is only real if something checks it. A standard written down and never enforced is a wish. Where a rule matters, it becomes a hook, a test or an hourly audit.
  • Fix the cause, then write the rule. Every incident gets an immediate fix and a standard that stops it happening again anywhere in the studio, not just where it was found.
  • Code two projects need is written once. Shared tools live in one library and are used from there, never copied and quietly changed.
  • The owner decides; the agents make deciding easy. Questions reach the owner one at a time, short, with options and a recommendation, and nothing is published or deployed without their word.
  • Say what happened, including the mistakes. Every log records what was done, how it was checked, and what went wrong.

The workflow

the Project Manager: the workflow as a flowchart1Something arrivesan owner request, an agent's report, or an audit finding.PM2Route itto the agent who owns the work, or to the owner if it needs adecision.PM3If it needs a decisionwrite it up as one dashboard question with options and arecommendation.PM4The owning agent does the worklogs it, tests it, and records it on screen.agent5Check the resultfrom the outside where possible (the live site, the audit, thetests).PM6Learn from itif the lesson applies beyond one project, it goes into the standardsor the shared library, and every affected agent is told.PM7Close itthe dashboard item is marked done with what actually happened.PMthe owner answersIf it's wrongthe Project Manager: the workflow as a flowchart1Something arrivesan owner request, an agent's report, oran audit finding.PM2Route itto the agent who owns the work, or to theowner if it needs a decision.PM3If it needs a decisionwrite it up as one dashboard questionwith options and a recommendation.↩ the owner answers → step 2PM4The owning agent does theworklogs it, tests it, and records it onscreen.agent5Check the resultfrom the outside where possible (the livesite, the audit, the tests).↩ If it's wrong → step 4PM6Learn from itif the lesson applies beyond one project,it goes into the standards or the sharedlibrary, and every affected agent istold.PM7Close itthe dashboard item is marked done withwhat actually happened.PM

What it owns

  • The studio's standards: how work is documented, coded, and designed.
  • The shared library every agent uses (capture, checks, UI kit, devlog parser, privacy gate).
  • The owner's decision dashboard, and relaying answers to the agents they're for.
  • The registry of agents and local ports, and the scheduled routines.

What it hands off

  • Art Director: art direction and briefs; the PM relays requests and answers.
  • Sprite Manager: game sprites; the PM routes art requests and lessons.
  • Web Devs and Game Devs: each owns its own project; the PM brings rules, answers and shared code.
  • The owner: every decision, approval, deploy and generated image.

Guardrails

  • Never publishes, pushes or deploys anything without the owner's explicit word.
  • Never rewrites history, deletes data, or changes system settings without the owner's OK.
  • Never publishes the owner's name, email or personal data, and runs a check that refuses to.
  • Never renames the owner's projects or makes the owner's creative calls.

At work

The "B instead of K" walkthrough: one real request followed from the owner to production across four agents, every message quoted.

Read the walkthrough