Testing a Fable-Style Agent Problem-Solving Plugin
alexcovo_eth · x · 2026-07-15
This post recaps an article where someone deconstructed Fable's problem-solving method into a plugin runnable by any model, conducting 159 agent runs for testing.
The core design has three components:
- fable-method: Provides the model with a structured problem-solving loop and sets hard failure thresholds.
- fable-loop: Completes the entire task while using adversarial validation agents to check the results.
- fable-judge: Treats all claims of "completed and passed tests" as unverified and independently re-runs them.
Test results include:
- Haiku improved from 0/4 to 4/4 in identifying "failing tests".
- Sonnet + method matched Fable's own performance on a five-step research task, scoring 10/10.
- The judge boosted Haiku's ability to detect fake completion reports from 3/5 to 5/5.
The author believes its true value lies in "knowing when not to apply it": for basic tasks and already highly capable models, this method yields little to no extra benefit.
More from coding & agent
- Codex helps build Valdiluce, an open-world game with climbing, gliding and gondolas — Dimillian · 2026-07-22
- HeyGen adds a media-sourcing skill for coding agents with 75k images and 10k tracks — HeyGen · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- LangSmith adds tracing for Pipecat, LiveKit, OpenAI Realtime, and Gemini Live — LangChain · 2026-07-22
- An MCP server signs every AI agent tool call into a verifiable Merkle chain — Funky_Chicken_22 · 2026-07-22
- Annotated transcript of a Claude Code team interview is now available — trq212 · 2026-07-22