The problem
Running several AI agents without an operating model produces the same failures over and over: work gets routed by hand, every session reloads the same context from scratch, handoffs happen in prose, and nobody can say afterward which agent did what or why. The agents were capable. The system around them was the gap.
The operating model
I designed and operate a lab where multiple agents (Claude Code, Codex, Hermes, and local models) work under one set of conventions.
- Routing. Interactive workspace sessions are orchestrated by one agent; ambient work arriving through chat channels is orchestrated by another. Which agent leads depends on where the work enters, and the rule is written down.
- Shared memory. A cross-agent memory tier, Open Brain, means I explain context once. Any agent can retrieve a decision, a constraint, or a prior incident without me re-telling it. A shared protocol file defines a single source of truth for each class of fact, so the stores do not drift apart.
- Context discipline. Large tool outputs route through an open-source MCP guardrail, context-mode, which indexes them and returns searchable summaries instead of flooding the conversation window. I integrate and operate it; I did not author it.
- Receipts. Work leaves a trail: an operational review queue, structured task and event records, and commit conventions that make delegated work auditable after the fact.
Safety gates
Autonomy is granted by class of action, not by trust in the agent. Reversible, in-scope work runs without asking. Deletes, restarts of shared services, credential changes, and anything outward-facing require human approval. Enforcement lives in hooks and permission rules rather than in prose instructions.
What it runs on
Python and SQLite carry the orchestration and record-keeping, Git carries the audit trail, and local models handle routing decisions that do not need a frontier model. The workload counts it produces describe my lab, not customer adoption, and the underlying commercial and open-source tools are integrations rather than products I authored.