'External Intelligence': Failure Boundaries Left by Prior AI Lifted Correct Completions from 1/3 to 3/3
Powerful_Creme2224 · reddit · 2026-08-22
While running a real SOL task on Codex, the author saw 48.35M input tokens with a 98.7% cache hit and only a 1% weekly quota move — though he explicitly avoids claiming causality.
His real focus is External Intelligence: keeping past failures, decision rationale, stop conditions, restart points and discovered boundaries outside the chat (e.g., in Obsidian), so a later AI retrieves only the relevant pieces — e.g., "this completion judgment failed before" or "missing this boundary causes false completion."
He ran 3 matched pairs (baseline vs with External Intelligence):
- OpenClaw: baseline hit a hidden countercase (false completion); EI side passed.
- VS Code: same pattern — baseline false completion, EI side passed.
- AWS CDK: both correctly BLOCKED; EI side was actually slower.
Correct completions went 1/3 → 3/3; input tokens dropped while command executions rose. His narrow claim: knowledge left by previous AI work changed later decisions in some matched cases — not that it's always faster or cheaper. He argues stronger models don't automatically accumulate judgment about when to stop or which failures should survive, and uses this workflow for OSS fixes (21 direct commits so far).
More from coding & agent
- Liquid AI's d1-ultra decides instead of writing, charging input-only $0.042 per million tokens — JosephJacks_ · 2026-10-16
- Full course: build a playable Unity fishing game with Claude Opus 5.5 in six steps — Roger_M_Taylor · 2026-10-16
- Marc Lou praises PostShiba, a cute indie alternative to AWS SES for email infra — marclou · 2026-10-16
- RExBench: best coding agent completes only 33% of autonomous research extensions — najoungkim · 2026-10-16
- Wiring Pstack skills from a Grok bot into Hermes agent loops — alexcovo_eth · 2026-10-16
- past.dev launches long-term memory API for AI agents, tops BEAM public benchmark — thisiskp_ · 2026-10-16