Skip to content

Claude/agentic harnesses exploration 2qvpye - #148

Open
yap-r wants to merge 9 commits into
dsondak:masterfrom
yap-r:claude/agentic-harnesses-exploration-2qvpye
Open

yap-r wants to merge 9 commits into
dsondak:masterfrom
yap-r:claude/agentic-harnesses-exploration-2qvpye

Conversation

@yap-r

@yap-r yap-r commented Aug 27, 2026

Copy link
Copy Markdown

No description provided.

yap-r and others added 9 commits September 12, 2019 12:45
A runnable harness built up one concern at a time, so each addition can be
read as a diff against the previous stage:

  stage1  the bare agentic loop, one bash tool
  stage2  tool registry, parallel execution, is_error results, truncation
  stage3  workspace sandbox, permission engine, JSONL audit log
  stage4  token accounting, prompt caching, eviction vs. compaction
  stage5  subagents, streaming, pre/post tool hooks

The tool surface stops changing after stage 2; stages 3-5 import it unchanged
and add layers around it, which is the actual shape of harness design.

Includes a mock model so every stage runs without an API key, plus
EXPERIMENTS.md with graded exercises (four of which need no key).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FXoBTEnyg15WzmZxwjuQF6
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants