Self-Evolving Agent Harness: +34.7 Points Over Frozen Baseline Without Any Training
Prompt Engineering · youtube · 2026-11-15
The Prompt Engineering channel published a tutorial on building a self-improving agent harness without training: the agent starts with only shell, read, and write tools, then grows its own tools, lifecycle hooks, and memory while working, built on the Gemini Interactions API with mid-task tool swapping.
Key content:
- Three-zone architecture, createtool with a unit-test gate
- Lifecycle hooks (sessionstart, beforedone) and memory
- Hidden grader and secret broker design
- Three benchmark rounds vs a frozen harness: web apps, silent bug-report videos, a custom binary format; HARNESS-evolved hits 77.8, +34.7 pts over baseline
- Four rules for when a self-evolving harness pays off and when it bloats itself
Sponsored by Fish Audio.
More from coding & agent
- DeepMind's AlphaProof Code formally verifies Signal crypto with 6M lines of Lean in 4 weeks — dl_weekly · 2026-11-15
- Raindrop lets coding agents create accounts before humans claim them, with write-only keys — New-Tailor4071 · 2026-11-15
- Write the eval runner before the agent: evals-first architecture with Ports and mock adapters — hwchase17 · 2026-11-15
- How do you evaluate agents whose skill library keeps evolving between test runs? — tahaan · 2026-11-15
- One lint fix burned 50M tokens: agent plugin nukes Codex Pro quota in 12 hours — generativist · 2026-11-15
- Developer builds a due diligence AI agent that screens VAT, LEI and sanctions lists — andrejusb · 2026-11-15