How do you inject instructions mid-run? Dev struggles with stage summaries agents don't trust
tinker_bob · reddit · 2026-10-28
A developer building a staged coding pipeline (localisation → solving → testing) injects per-stage summaries as tool results to shrink context and cut costs — but the agent sometimes disbelieves the summary and redoes completed stages. He's weighing three injection approaches: a separate assistant message framed as "what YOU did," an assistant message framed as "what I did," or a user message recounting progress. No conclusion yet; the thread surfaces a real agent-engineering pain point around trusting compressed context mid-run.
More from coding & agent
- Judgment Labs launches publicly: monitoring and eval layer for long-horizon agents — testingcatalog · 2026-10-28
- OSU NLP Group Wins Three Paper Awards Same Day at EMNLP 2026 and NeurIPS Workshop — hhsun1 · 2026-10-28
- Replit launches Cloud Agents: non-engineers ship reviewed PRs from natural language — Jeff_Burke14 · 2026-10-28
- Sequoia and OpenAI host $100K poker bot tournament where agents build bots in 60 minutes — polynoamial · 2026-10-28
- PPCC26 demos praised: real orgs like NFL, real code live in VS Code, no fake companies — donasarkar · 2026-10-28
- Five principles for curating LLM agent evals: diverse, realistic, hard, discriminative, high-quality — cwolferesearch · 2026-10-28