01
Problem
Agent demos often hide the runtime decisions that determine reliability: what tools execute, what context survives, which permissions apply and whether a change improves behavior.
02
Role
Sole designer and engineer responsible for runtime architecture, safety boundaries, evaluation workflow and implementation.
03
Architecture
- 01
Provider-compatible streaming tool loop - 02
Episode, working and tool-memory context layers - 03
MCP stdio client, subagents and persistent sessions - 04
Replay-based skill candidate evaluation
04
Key contributions
- Implemented five permission modes plus read-before-write and modification-time conflict checks.
- Designed context folding that preserves actionable memory without unbounded prompt growth.
- Built an add–merge–discard skill evolution flow with rule and LLM judges, replay sets and champion selection.
- Tracked session tokens, cost and tool events for inspectable agent runs.
05
Evaluation
- Replay sets compare candidate skills against the current champion using deterministic rules and optional model judging.
- Runtime checks cover permissions, conflicts, tool events and context behavior.
06
Technology
07