Skip to content

Releases: 0Smallcat0/otto

Release list

v1.0.0 — Measured agent operability

v1.0.0 — Measured agent operability Latest
Latest

Choose a tag to compare

View all tags
@0Smallcat0 0Smallcat0 released this 10 Jul 10:58
v1.0.0
e81aca7

Otto's first complete release: a local, AI-operated financial terminal with a measured operator surface.

A real agent operating Otto

Measured agent operability

The claim "an AI can operate it" now ships with a benchmark (evals/): 20 plain-language tasks run by a real headless agent in hermetic sandboxes, graded programmatically (terminal state, produced artifacts, refusal-with-state-unchanged) — no LLM judge.

Model Tasks Passed Success rate Avg turns
claude-sonnet-5 20 20 100% 6.5
claude-haiku-4-5 20 19 95% 6.8

Full results and limitations: evals/EVAL.md. A smoke mode proves every graded check starts red on fresh state, so tasks cannot pass vacuously.

Highlights

  • Agent-native architecture — one typed contract (113 actions) is the single source of truth; the MCP tool surface, UI capability catalog, and eval suite all derive from it. Write-up: ARCHITECTURE.md + ADR-0002/0003/0004.
  • Structural safety — no live execution paths exist; secrets never transit agent tools; destructive actions need typed confirmation flags; safety tasks in the benchmark regression-test refusal.
  • Honest quant research rails — closed candles with next-open fills (lookahead guard), explicit fee/slippage economics, walk-forward validation with engine-issued consistency verdicts, overfitting flags in plain words. Methodology study: docs/research/.
  • 475 tests on Windows + Linux CI, plus Playwright e2e and the token-free eval smoke mode.

Notes

  • Results above are single-run-per-task (no retries); the eval report says so explicitly.
  • Otto is a clean-room rebuild from observed workflow shape — no third-party code, assets, or branding. The observation corpus stays unpublished by design.

Full changelog: CHANGELOG.md

Assets 2
Loading