Skip to content

Agentagon 0.1.3

Agentagon now helps you establish a quality baseline with Init, improve it within an agreed budget with Fix, and inspect or rerun measurements in the dashboard.

Changes

  • Add Init onboarding that recommends a useful first measurement, reuses existing evaluations and keeps detailed definitions in saved evidence. Deterministic checks remain distinct from measurements of task quality.
  • Save behavior and scoring definitions, versioned evaluators and immutable baselines. Rerun compatible committed application code while preserving earlier measurements and keeping recent-trace scores separate.
  • Add bounded Omni optimization with GEPA and Agentagon's native-host AutoResearch and Meta-Harness adapters. Repeated trials count against the evaluation budget, and final verification retains its reserved capacity. Agentagon's optimizer runtime manages bounded evaluation and parallel stages; GEPA search uses the published gepa==0.1.4 dependency.
  • Add dashboard baseline history, explicit reruns, settings and report exports. These journey and dashboard changes are in #8.
  • Reduce retained experiment storage by sharing content-addressed evidence and releasing eligible finished worktrees. Preserve recovery records, user edits and older inline artifacts. Export fix reports explicitly and retain telemetry events through brief write contention. #7
  • Move review and delivery into the primary journeys, support reviewed unmeasured patches and local delivery, and require per-request approval for optional Intelligence by default. #6

Compatibility and upgrade

Requires Python 3.12+ on macOS or Linux. This pre-1.0 release includes changes to the exposed skills: use Audit for change review and Fix for delivery instead of the former standalone ag:review and ag:ship skills. Init and Eval are available for baseline and evaluator work. Existing Intelligence installations require approval for each outgoing request unless full access is explicitly configured.

After the package is published:

pipx install agentagon==0.1.3
# For an existing installation:
pipx upgrade agentagon
agentagon install --host codex
# Or: agentagon install --host claude-code

Restart the coding host after registration to load the updated skills. Earlier release artifacts and frozen evidence remain unchanged. New baseline and optimizer records do not convert old runs into fresh measurements. Resume unfinished optimizer runs from unreleased GitHub-dependent builds using their original installation; 0.1.3 rejects resuming a different optimizer runtime so it cannot mix search implementations in one saved run.

Validation

Release artifacts are agentagon-0.1.3-py3-none-any.whl and agentagon-0.1.3.tar.gz. The release workflow adds the exact source revision to its staged notes and records both artifact hashes in SHA256SUMS.

Local validation on 2026-09-15 used macOS arm64, Python 3.13.2, Node.js 22.14.0 and the published GEPA 0.1.4 wheel:

  • Non-browser suite: 873 passed, 9 optional integration checks skipped.
  • Final optimizer and runtime checks after adding the resume guard: 14 passed. These include the new guard and concurrent evaluation limits.
  • After the runtime module rename and documentation cleanup, focused runtime, optimizer, optimization workflow, release and website checks: 43 passed.
  • Dashboard browser checks with Playwright 1.57.0 and Chromium build 1200: 14 passed at mobile and desktop widths.
  • Ruff lint and format checks, shipped JavaScript syntax and installer shell syntax passed.
  • Wheel and source archive passed strict Twine metadata checks. Wheel inspection confirmed the optimizer runtime, license notices and absence of direct URL dependencies or a copied GEPA package.
  • A fresh environment outside the checkout installed the wheel and its PyPI dependencies. Version, resources, the local audit example, dependency consistency and a bounded offline GEPA search passed.
  • The extracted source archive built and checked 43 documentation pages; all 5 source-package website tests passed.

Known limitations

Measurements require runnable application code, suitable evaluators and authorized execution and model access. Missing measurements remain unknown; an unmeasured patch does not establish a verified improvement. Coding-agent judgments are labeled separately from executable or API-backed checks.

Native-host registration, live SSH/E2B execution and live telemetry ingestion are separate integration checks; the ordinary suite and offline installation example do not establish those integrations.