The momentum prior: how early tokens lock agents onto a trajectory — escape by fresh context
gerardsans · x · 2026-09-09
Developer gerardsans names a failure mode: the "momentum prior." Sitting next to context rot, it describes how prefill and early tokens set a prior like a truck going downhill — the residual stream reinforces the original task direction until later instructions are just a tap on the wheel, not a new road. A quoted Japanese tweet illustrates the symptom: running coder and reviewer sessions in a loop over gh issues makes the reviewer drill into ever-smaller problems, endlessly spawning new issues and even trying to PR others' repos — an agent-turned-token-consumption-machine.
Core argument: attention is not a magical spotlight but a routing mechanism with hard capacity limits; quadratic complexity buys the residual stream but locks in "momentum hijacking" failure modes that scaling alone can't fix. Escape: start a fresh context rather than fight a locked-in trajectory.
Related event: 'Momentum Prior' and Context Rot: Why You Should Restart Sessions(3 posts)→
More from coding & agent
- Sol61 flunks instruction-following: skips mandated research even when explicitly told — lucasmeijer · 2026-10-26
- Ben Bajarin: Frontier AI is becoming the enterprise cyber-defense stack — BenBajarin · 2026-10-26
- Building AI bots on raw APIs? Devs say models are 'dumb without the harness' — madeofAjala · 2026-10-26
- HarnessMM cuts agent tokens per task by up to 88% while raising success rates — omarsar0 · 2026-10-26
- Microsoft demos intent-classifying agents built with Copilot App and Foundry, deployable across Teams and Outlook — marlene_zw · 2026-10-26
- Looking for a wearable mic with best battery life that feeds recordings to personal agents via MCP — altryne · 2026-10-26