[new-rules] Reworking the agent harness based on Anthropic + Google papers — feedback wanted #999
mhenke
started this conversation in
Show and tell
Replies: 1 comment
|
https://github.com/alvinunreal/oh-my-opencode-slim/tree/omos/new-rules To try it locally: Then point OpenCode at the branch. In OpenCode, run: Restart OpenCode after that — it loads plugins on startup, so the restart is what actually swaps in the new code. If you don't have the upstream remote yet: |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Two papers changed how I think about agent harnesses.
Anthropic (The new rules of context engineering for Claude 5 generation models, Aug 2026) cut 80% of Claude Code's system prompt with no eval loss. They swapped rules for judgment, examples for interfaces, upfront context for progressive disclosure. Google (The New SDLC with Vibe Coding, May 2026) frames "vibe coding" vs. "agentic engineering" as a harness problem, not a model problem.
This branch applies those ideas to OMOS.
What changed
src/agents/orchestrator.ts): simplified from rigid rules to a routing threshold with judgment guidance.SKILL.mdfiles pointing toreferences/full-guide.md. Agents load deep content only when needed.src/evals/): assertion-based framework that verifies progressive disclosure actually works end-to-end.What I'm deferring
/conductortoggle or bounded-autonomy loop.What I'm asking
Branch:
omos/new-rulesKey files:
src/agents/orchestrator.ts,src/evals/,src/observability.ts,src/skills/*/SKILL.md+references/full-guide.mdFor developers: Eval CLI:
bun run eval. Add cases insrc/evals/orchestrator-routing/eval.json. Cloned deps at.slim/clonedeps/repos/are read-only. this is manual. you paste in the prompt, then copy the results.All reactions