Actspan Start with Actspan
Actspan is an AI agent control plane that connects versioned configuration, policies, evaluations, and runtime evidence. Use that connection to understand what changed, which configuration ran, and which checks support the result.
Start with the CLI: define behavior in YAML, run it from your terminal, evaluate changes, and inspect the recorded results. Connect the shared portal for team review, configured release requirements, and linked evidence.
Start with a small prompt to check your installation. Then choose the governed refund to see an action proceed, wait for review or be refused. The research workflow introduces two agents and evaluation of their handoff.
Install → check a small prompt → run a governed refund → inspect its receipt
The public download provides five native examples for CLI 0.2.0; no repository account or partner handoff is needed. The current native download supports Apple Silicon macOS. Optional adapters, the Python SDK and portal access are supplied separately for design partners.
Start here
| Your next step | What you will learn |
|---|---|
| Why Actspan | Assess the control problem, alternatives, and measures of value |
| Install Actspan | Set up the CLI and credentials |
| Run the quickstart | Create a two-file no-tool prompt, check it offline, then make one live request |
| Try a governed refund | Inspect automatic payment, local review, refusal and a persisted receipt |
| Understand the core concepts | Map the released research example to prompts, agents, workflows and runs |
| Understand the research and summary workflow | See how the research result reaches the summarizer |
| Change the summary | Edit its instructions and evaluate the effect |
| Browse OpenAI examples | Choose a workflow and follow its configuration, controls, and evidence |
The local path uses the public CLI and examples, plus your provider key for live calls. The default bundled OpenAI path needs no host Python, portal account or Kubernetes. Optional isolated provider and framework packages have additional installation requirements described in their guides.
A configuration-first workflow
Define → Validate → Preview → Run → Inspect → Evaluate
Validation checks configuration. Preview estimates what will run without calling the model. A run executes the selected target. Evaluation checks its behavior against examples and explicit rules.
When configuration declares supported long-lived provider resources, apply
reconciles them. run can also ensure those resources before execution. Applying
resources and executing an agent are distinct operations.
Choose a task
| Task | Guide |
|---|---|
| Change a model or prompt | Change a workspace |
| Add retrieval or another external capability | Connect an MCP server |
| Check output quality | Evaluate behavior |
| Find what a change affects | Review change impact |
| Test a supported policy boundary | Verify controls |
| Investigate a run | Inspect local evidence |
| Share results with a team | Connect the portal |
Understand what each result means
| Result | Meaning |
|---|---|
| Configuration valid | Its structure and references pass validation |
| Run succeeded | Execution completed; answer quality needs evaluation |
| Evaluation passed | The declared cases and checks passed |
| Verification passed | The specifically reported control checks passed |
| Missing or unsupported | Coverage or evidence is absent; this is not a pass |
Individual concept pages describe current support and limitations. In particular, some agent runtime fields are declared but not fully enforced, and vector-backed agent memory remains future work. See Agents before depending on those controls.
Use the reference when you need it
Start with CLI workflow, Python SDK, or a provider guide such as OpenAI and Anthropic. Detailed field references remain linked from those pages.