Skip to content
Saafir HaleSecurity & InfrastructureAtlanta, GA
← Software & Apps

Director Pattern.

The control plane for an autonomous agent hierarchy: earned trust, a hard budget, and a grader that is never the model it grades.

Personal build
About

Director Pattern sits on top of AI Orchestra. Cloud models act as managers that plan work and review it, and the local models on the LAN act as workers that fetch, draft, analyse and summarise. It runs on its own every six hours. I built the control plane first, on purpose: a hierarchy of agents that cannot score its own work or stop its own spending is a liability, not a system.

Autonomy is earned per branch and per kind of action. Anything that leaves the machine starts gated and needs a human approval, moves to trial after 5 clean instances and to auto after 10 more, and drops a full level the moment one is rejected or fails verification. Some actions, like signing anything or touching a payment method, never promote at all. Trust widens how often a branch may act, never how large one action can be.

Spending is refused at the point of booking, not advised against. Every call is priced before it runs, each branch has a budget envelope and a per-action ceiling, and a kill switch is checked before every spend and every outward action. Releasing it takes a literal confirmation phrase, so a retry wrapper or an agent that finds the method cannot clear it.

Work is graded by a model from a different family than the one that produced it, which is checked when the code loads. Success is never a model's opinion: each goal names a measured fact, and those metrics are written by separate collectors rather than by the agent being judged. An agent that can write its own scoreboard has no gradient, only an incentive.

The live runs shaped the code more than the design did. When the primary cloud model kept returning 503s, the system gained retries, a priced fallback model, and rate-limit backoff, and a manager failing became a recorded lesson instead of a crashed run. When a small local model merged three useful summaries into "nothing relevant", empty partial results were dropped before the merge and the merge became a union. 154 tests cover it.

Built with
PythonSQLiteFastAPIOllamaClaude / Gemini APIspytestTask Scheduler
NextQuantitative Trading Research →