Context Management
Context engineering for long-horizon, multi-step agents.
My agent work uses context management to keep long tool-use trajectories relevant and auditable.
- Progressive loading separates skill summaries from detailed execution knowledge.
- Task-relevant evidence and reusable lessons enter the context when they support the current decision.
- Clear boundaries distinguish agent-visible observations from service-side repairs and hidden reference information.
These ideas inform the FIRE harness and my ongoing Agentic Search data and training design.