Quick Start

After installing and connecting, list the available workflows and choose agentic-engineering, the default starter. Its short alias is engineering:

Use the engineering workflow to investigate and fix the failing test.

In Claude, you can also open /statewright, select the workflow and choose Start workflow. In Codex, ask for statewright_load_workflow with name: "agentic-engineering". Follow the actual state and transition instructions returned by the tool.

The same starter is available through Cursor, OpenCode, Pi, OMX, and other MCP hosts. Use your integration's controls to load it; the host determines which tool and model controls are available.

Understand before changing

The diagnose phase reads the relevant code and captures your outcome and acceptance criteria. Baseline reproduces a bug, characterizes behavior before a refactor, or defines how to verify a new feature. Neither phase exposes source editing. Baseline can run bounded local checks where the host supports them.

Implement, validate and review

Implement enables a focused change and regression coverage. Validate runs the declared checks against that change. Review takes a separate critical look at the result before handoff summarizes what changed, what was checked and what remains uncertain. It does not automatically commit or deploy.

A failure gets one evidence-led repair round, followed by fresh validation and review. If that fails, the workflow ends as blocked, with a useful next action rather than an indefinite retry loop. A pending native permission prompt is not a denial; answer it in your client to continue.

Investigations and no-change work

You can also ask:

Use the engineering workflow to investigate this error. Do not edit anything.
Use the engineering workflow to add input validation to this form.
Use the engineering workflow to refactor this parser without changing behavior.

Read-only investigations and already-correct behavior skip implementation but retain evidence, validation and review. The workflow uses your selected model; it does not require a premium model or the larger V2 profile's configuration. Supported clients apply phase-specific effort; other hosts keep their settings. Model changes per phase are optional configuration for supported clients, not a requirement of this starter. Any demo showing Haiku, Sonnet, and Opus illustrates a configured workflow; it does not report measured token use or savings.

Completion requires recorded validation evidence, but that record is still the agent's report—not independent proof. Check the actual commands, results and limitations in its handoff. A review phase is not necessarily a separate reviewer. Native permissions and Statewright approval gates remain in effect.

Every existing account receives this starter alongside its current workflows, and new accounts start with it. It is the default, independent of which workflow you edited most recently. You can delete the old bugfix workflow. Customized workflows and run history are retained; an existing canonical definition is not overwritten or re-enabled if you explicitly disabled it. If you already have a workflow named engineering, that exact name takes precedence; use agentic-engineering to select the starter unambiguously.

View the Run

Go to statewright.ai/runs to see transitions with timestamps, rationale, and observed tool calls per phase. The rationale strings are stored with events; they are the agent's stated reasons, not a record of its private reasoning or proof that every host tool call was intercepted.

For the theory behind all this, see Core Concepts. To build your own workflow, Create Your Own.