Govern every step of a complex workflow.

THE CONTROL KNOBS

Every step carries its full control set — tighten-only.

A Workflow is a DAG of steps. Each step composes tighten-only over your org/team acceptance policy — it can add a rule or raise a bar, never remove an org rule or lower one. So you can safely delegate per-step control to a team, or to an external system.

• policy_ref — the stored acceptance policy that governs this step
• bands — per-step decision thresholds (ship / route_to_fix / quarantine / human_approval_below)
• hard_rules — deterministic gates (forbidden/required term, PII, regex, citation-required, min-score, max-flaw-severity)
• requirements — preconditions to ENTER the step (required input, upstream decision/score, manual sign-off)
• approval — a human-approval gate — required? which roles? SLA? escalation? on-timeout?
• routing — conditional branches (on_decision), terminal decisions (stop_on), bounded iterate-on-fix loops (max_visits)
• retry / timeout_secs / cost_cap_usd — execution guards
• provider / model — per-step routing hints

THE CONNECT LOOP

Two calls in a loop: plan, then evaluate.

1. Fork a governed workflow from 1000+ presets in one call: POST /api/v1/workflows/from-template (browse GET /api/v1/workflows/verticals / …/templates).
2. plan — pass the run state you have so far; get the next runnable steps plus each step's resolved governance. The driver enforces the DAG, conditional routing (branches + bounded loops), and requirements, so you never reimplement the governance.
3. evaluate — POST /api/v1/workflows/{id}/steps/{step_id}/evaluate grades each step and returns the governed decision.
4. loop — fold the decisions back into completed and call plan again until done.

POST /api/v1/workflows/{id}/plan
{ "completed": { "draft": { "decision": "ship", "score": 0.9 } },
  "inputs": { "query": "..." }, "signoffs": { "legal_ok": true } }

-> { "ready":   [ { "step_id": "review",
                    "controls": { "bands": {...}, "approval": {...}, "policy_chain": [...] } } ],
     "blocked": [ { "step_id": "...", "unmet_requirements": [...] } ],
     "suppressed": [ "..." ],        // a branch the routing did NOT take
     "done": false, "terminal": null }

DROP-IN SDK

Zero-dependency clients — Python & TypeScript.

Copy one file (sdk/workflow/) and the whole plan → run → evaluate loop is one call.

from seaotter_workflow import WorkflowClient
otter = WorkflowClient(base_url="https://api.seaotter.ai", api_key="sk-otter-...")
otter.from_template(vertical="comic-drama-video-mantur-ai", new_id="my-wf")

def run_step(step, controls):
    return {"artifact_text": my_orchestrator.run(step["step_id"])}

result = otter.run("my-wf", run_step, inputs={"query": "..."}, grade_live=True)
print(result["terminal"])     # { "decision": "ship" | "block" | ... }

GO DEEPER

Machine-readable contract.

• /llms.txt — the agent entry point (step 6a is the workflow connect loop).
• OpenAPI spec — full schemas, tags workflows + workflow-templates.
• Agent-native contract · the OtterScore eval loop these step grades run on.