Drew Breunig walks through a problem familiar to anyone running a coding agent for hours: its accumulated history can keep a bad idea alive. His AGNTCon talk uses DSPy, a Python framework for programming model workflows, to show how the surrounding software can manage that history.

The six tactics are concrete:

  • Augment: bring in relevant documents and instructions when needed.
  • Choose tools: expose a task-sized subset of the available tools.
  • Quarantine: give a self-contained subtask to an agent with a fresh conversation.
  • Prune: remove history that has stopped helping.
  • Offload: save durable information in files or a database.
  • Summarize: replace a long exchange with a shorter briefing, potentially after selecting its relevant parts.

Breunig calls repeated pursuit of an exhausted hypothesis “context inertia.” He describes systems that clear the conversation while preserving work in the environment. His closing concern is longer-lived memory: skills can accumulate narrow fixes until the instructions conflict.

Caveat: His roughly 200,000-token warning point is a personal heuristic. The short code examples illustrate designs; they do not establish a universal reset policy.

What the thread adds

The 52-comment thread on Hacker News adds counterexamples and implementation choices:

  • ethanj8011 — challenges the reset threshold from coding experience: “Without that context I find the models perform badly.” They report little change until around 700,000–800,000 tokens, an anecdote that argues for testing the workload.
  • DenisM — reports checkpointing code, design and intent, then asking another agent to validate the approach before resuming. “Unexpectedly, design doc and an execution plan are now outcomes of coding process, not inputs to it.”
  • infogulch — points to a hierarchical memory design: increasingly compressed older messages, plus a “zoom” tool to recover detail. It offers a way to make removed context retrievable.
  • agentdev001 — raises the economics of pruning: changing an earlier turn may invalidate reuse of later cached input. They explicitly say they have not tested that understanding and ask for comparisons of provider-hosted and custom compaction.
  • losvedir and zby — question whether the Pokémon tea mistake diagnoses a context problem or a wrong association learned during training. WarmWash replies: “Once pulled it added “iced tea” to its to-do list, which gets put into context.” That matches Breunig’s account of how the mistake persisted. zby’s remaining objection is about the remedy: “cleaning the context would not fix it - because it would be poisoned on next round”. The example describes persistence; it leaves open whether a reset alone prevents the mistake from recurring.

HN handles are pseudonymous. There are no public per-comment scores; ordering is HN’s own ranking. This is a slice of the thread, not a consensus.