What is Staso?

Staso is a Python SDK and platform for AI agent observability, guardrails, and automated failure diagnosis — built for teams running Anthropic, OpenAI, Claude Code, or Codex agents in production.

What it does

Observe. Staso traces every agent run end-to-end: LLM calls, tool invocations, token usage, latency, errors, and intermediate state. One patch_anthropic() or patch_openai() call captures model requests, responses, and streaming output. Decorators like @st.agent and @st.tool add structured spans to your own code. See LLM tracing and integrations.

Guard. Staso Guard evaluates tool calls before they execute. Block destructive operations, modify unsafe arguments, or escalate to a human — enforced from a single rules surface. See Guard overview.

Heal. Point Heal at any trace — a failure, a regression, a slow run, a bad eval result — and it diagnoses the root cause from the span tree and suggests a fix. For CLI agents (Claude Code, Codex) the fix can land as a PR. See Heal overview.

Who it's for

Python teams shipping AI agents to production. Staso plugs in without rewrites. The SDK is Python 3.11+ and supports sync, async, and streaming responses.

How it fits together

1. Your agent runs — a Python process or CLI invocation.
2. The SDK traces it — spans flow to Staso in the background.
3. Guard checks tool calls in-flight — blocks, modifies, or allows.
4. When something looks off — a failure, a regression, a latency spike, a bad eval — Heal diagnoses the trace and proposes a fix.
5. You review and merge — or let the agent apply it for CLI targets.

Each piece is independent. Start with tracing, add guards when you need enforcement, turn on Heal when you want diagnosis automated.

Next

Read the Quickstart to get a trace on your dashboard in under five minutes, then learn the core concepts.