The tool-vs-context split breaks with multi-step tool chains — split by resolve time instead
JaseciLabs · reddit · 2026-09-30
The JaseciLabs team shares a framework from building agentic support for Jac: the "do → tool, know → context" heuristic breaks down in multi-step ReAct chains. The real split is whether a value can be resolved before reasoning starts or at a specific moment during it. Pre-resolving request-time-known data (policies, account tiers) into static context beats exposing it as a tool — saving a round trip and removing the failure mode where the model forgets to call it; live values (inventory, pricing) need real tools. They ask whether anyone encodes this as an explicit per-dependency "resolve-time" rule.
More from coding & agent
- OpenAI DevDay 2026: Dots always-on agents, GPT-6.1 Sol at 1/5 token price, cloud Codex — TheMoonMidas · 2026-09-30
- One of a user's 40+ OpenAI Codex agents mysteriously tried to install Cursor — baditaflorin · 2026-09-30
- Runway shows how to turn one product photo into a full social ad campaign — tlakomy · 2026-09-30
- Google deploys agentic pre-submit scanning across millions of lines of infra code — 92% precision, 3% false positives — dl_weekly · 2026-09-30
- How IBM built Torch Spyre on PyTorch's CRCR to scale accelerator CI testing — PyTorch · 2026-09-30
- How much should agents post in team chat? Slack is just the coordination layer — Al_Grigor · 2026-09-30