Summary
IndyDevDan builds a self-compacting AI coding agent using the Pi coding agent, tackling a persistent problem in long-running agentic workflows: the context window filling up and degrading performance. Rather than relying on the default auto-compaction found in tools like Claude Code and Codex, this devlog shows how to give an agent its own dedicated tool to decide when to compact, with three configurable thresholds — notice, warning, and force-compaction.
The video walks through the full harness engineering process, starting with a written plan document, a custom “definition of done,” and strict grading rules that penalize agents for writing outside their designated working directories or reading files they shouldn’t touch. Dan then hands the same plan to three different agent harnesses running different models to compare how each one handles self-directed compaction in practice.
Aimed at engineers running long-horizon, multi-agent “swarm” systems where dozens to hundreds of agents work for hours unsupervised, the video argues that context management is an underused lever for controlling both cost and reliability. It’s a practical look at harness and prompt engineering techniques for making autonomous coding agents more context-aware and trustworthy over extended runs.
📺 Source: IndyDevDan · Published September 21, 2026
🏷️ Format: Hands On Build
