Before (the problem)
A strategic thinking partner is only useful if it has continuity — yet most AI
agents start each session blank, lose the thread between conversations, and run with
no defined model, tool inventory, or thresholds. An ungoverned agent that touches
strategy is exactly the risk the firm’s governance platform exists to control.
What we built
A memory-backed strategic Chief-of-Staff agent, governed under an approved CM-AI
baseline:
- Persistent memory of priorities, decisions, and open questions across sessions
- Four operating modes — reader (session start), writer (mid-session capture),
triager (inbox), synthesizer (weekly cross-session pattern review)
- Governed by an approved CM-AI baseline: defined model, tool inventory, KRI
thresholds, and a tested rollback
How it works
- At session start, Chloe reads working hypotheses, open questions, and recent
decisions to pick up the thread.
- During work, out-of-scope ideas are captured to the inbox instead of derailing.
- The inbox is triaged first thing each session.
- Weekly, Chloe synthesizes the last 7 days into updated working hypotheses.
- Every mode runs inside the baseline’s model + tool + threshold controls.
Outcomes
- 4 operating modes running across every session, with continuity preserved
- Operates inside the same four-artifact baseline the platform enforces for
clients (BASELINE, KRI_THRESHOLDS, ROLLBACK, CHANGE_LOG)
- Rollback rehearsed and passing — a governed production agent, not a demo
Stack & role
Claude · CM-AI baseline · Persistent file-backed memory. Built & operated in-house.
Timeline
One of three Phase 1 governance pilots; live and in daily use.
What it proves
The governance platform on its own dog food: a live strategic agent operating inside
the exact baseline-and-drift controls the firm builds for clients. The methodology is
demonstrated, not just described.
