agentleFS
Sign inStart free
IgorGanapolskyIgorGanapolsky/ThumbGate/.agents/skills/harness-evolution-ceiling/SKILL.md

harness-evolution-ceiling

Skill29 starsChanged 7 months agoMITView on GitHub

arXiv:2610.11655 diagnose-then-intervene rail: read failure composition (process vs content) to detect when harness evolution hits its ceiling. When content failures dominate, switch to weight training (LoRA/distillation) or model tier escalation instead of bloating prompts and AGENTS.md rules.

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    When process failures drop and content failures dominate ($C_f \ge 80\%$, $P_f \le 20\%$), declare Harness Ceiling Reached.

    line 17

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Harness Evolution Ceiling (arXiv:2610.11655 FORMAT Steal)

Goal

Apply the empirical law from Harness Evolution Hits a Ceiling: When Weight Training Should Begin (arXiv:2610.11655) to agent harnesses, task outcomes, and omp.sh:

  1. Label failed agent trajectories by the first signal that fires into Process Failures vs Content Failures.
  2. Evolve the harness (rules, gates, tool contracts) when process failures are active ($P_f \ge 20%$).
  3. When process failures drop and content failures dominate ($C_f \ge 80%$, $P_f \le 20%$), declare Harness Ceiling Reached. Freeze prompt rules and switch the lever to Weight Training (LoRA/distillation) or Model Tier Escalation (Route: named to frontier model).
  4. Partition accepted edits so security controls remain in the runtime harness while formatting/planning templates are isolated for weight adaptation.

Constraints

NEVERALWAYS
Bloat AGENTS.md / prompts with negative rules when failures are content/reasoning gapsRead failure composition first: classify into process vs content
Claim harness ceiling with $<3$ failures or when process failures are still high ($>20%$)Evolve harness (fix tool syntax, gates, loops) while process failures remain
Move security gates, checkout leases, or task-scope fences into model weightsKeep security controls in the runtime harness; train only templates & styles
Use unnamed models (Auto, Fast) when content ceiling is reachedEscalate to named frontier model (Route: named) or train LoRA adapter

Commands

# Audit harness ceiling rail status
npx thumbgate harness-ceiling --json

# Rail map
npx thumbgate harness-ceiling --map-only --json

# Dynamic failure composition evaluation
npx thumbgate harness-ceiling --evaluate --process=0 --content=8 --json

# Test suite
node --test tests/harness-ceiling.test.js

Reference

  • Paper: arXiv:2610.11655 (Tian et al., Oct 2026)
  • Doctor: scripts/harness-ceiling-doctor.js
  • Task outcomes: scripts/task-outcomes.js (classifyFailureType)
  • RRSI regularizer: scripts/rrsi-regularizer.js (evaluateHarnessCeiling, partitionAcceptedEdits)
  • omp coordination: scripts/omp-coordination-honesty.js (diagnoseFailureIntervention)
  • Documentation: docs/agents/harness-evolution-ceiling.md

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free