Why Your AI Agent Suddenly Gets Stupid And How to Fix It

OpenClaw agents degrade predictably past ~40% context — tighter SOUL.md and fewer skills fix most of it.

The "Dumb Zone" Is Context-Driven

The creators describe a recurring failure mode: once an agent's context window fills, it stops calling the right tools, "forgets everything," and "tells you to walk to a car wash." MiniMax 2.5 specifically gets "really stupid" once context exceeds 40%. This is the same agent, same model, same task — the drop is purely a context saturation problem, not a capability one.

Cut SOUL.md To 15–30 Lines

Jeff's SOUL.md was scrubbed from ~100 lines down to 28 lines, mostly ~6 lines of actual instruction. The default file shipped by OpenClaw is longer than that and performs worse. The rule the creators settled on: 15–30 lines maximum, with characters minimized. Coherence of outputs (e.g., news roundups where stories actually relate) jumped immediately after the cut.

Cap Skills At 7–10 For Cheap Models

Jeff runs MiniMax 2.5 and has news + web search only. Skills like Spotify, Discord player, GitHub, and weather were still installed from the default onboarding and were bleeding context (song titles, lyrics, JSON chatter from Spotify) into unrelated tasks. Recommended ceiling: 7–10 skills per agent on cheaper models; more than that makes them "dumber and dumber."

Stark (Opus 4.6) Proves The Pattern Repeats

Stark on Opus 4.6 produced a usable daily briefing, took initiative to flag MWC 2026 + Apple as trending "intent," and surfaced crypto + US macro without being told. But it refused to use Remotion for animations and the slides "look like crap." Fix in progress: write the presentation workflow as a skill and force it to show its steps for review.

Specialize, Don't Generalize

The team's working principle now: hyper-specialized agents doing one task well beat a general-purpose agent managing everything. Pair this with their prior videos on sub-agents and cron jobs for the long-term setup.