Connect, test, and release an agent.

Use Agent Studio or the public Runs API. Each run records the exact agent version it executes.

Connect one workflow

Create or import: Start with a cloud task, chat, or phone agent in Agent Studio. Retell exports can be imported as drafts; importing does not change the live version.
Open Agent Studio
Interfaces: Studio, REST, the TypeScript and Python SDKs, and MCP expose the same agent operations. Create a key in Studio → Settings → API keys to use https://api.cumuluslabs.io.
Documentation
Add a tool: Connect a customer HTTP action with its input/output schema and credential reference. Add the action to the flow and define which returned fields the agent can use.
Set the contract: Choose the model, instructions, input/output schemas, knowledge documents, and execution limits.

Test a pinned version

Studio Test publishes a runnable version without changing what is live. A test run suppresses customer lifecycle webhooks and event triggers. Connected tools still execute; use a test endpoint for actions that change external data.

Run your published cloud task

curl https://api.cumuluslabs.io/api/v1/runs \
  --header "Authorization: Bearer $CUMULUS_API_KEY" \
  --header "Content-Type: application/json" \
  --header "Idempotency-Key: $RUN_REQUEST_KEY" \
  --data '{
    "channel": "cloud",
    "agent_id": "YOUR_AGENT_ID",
    "version": 1,
    "mode": "task",
    "input": {},
    "test": true
  }'

Use your agent ID, published version, and input schema. Persist the request key and body; retry the same request with the same key.

Inspect the result and trace

Run status: GET /api/v1/runs/{run_id} returns execution state and the pinned publication. Admission is separate from completion.
Execution: Read /events for the timeline and /effects for tool actions and their receipts. Read /result for the task output and /usage for metered usage.

Compare before release

Compare a candidate publication with a baseline on existing runs. Evaluation reuses recorded tool receipts instead of sending external actions again; a missing receipt cannot produce a new external action. A Studio test and an evaluation replay have different tool behavior.

Create a comparison: POST /api/v1/comparisons with agent_id, baseline_version, candidate_version, case_run_ids, and rubric: {"id":"run-pairwise","version":1}. Cases must be completed runs of that agent. Review per-case verdicts before release.
Make the tested version live: Promote the reviewed publication in Studio. The API uses POST /api/v1/agents/{agent_id}/head with version and expected_head_version. Existing runs retain their admitted version.
Roll back: Move the live head to an earlier publication. Publications are immutable, so the tested version and its run history remain identifiable.

Choose the execution boundary

Cloud: Task and chat runs use the same public API. Background agents can start from schedules or events.
Phone: Bring your SIP trunk and numbers. Talos executes the conversation, records its run, and exposes signed lifecycle events; carrier connectivity remains part of your telephony setup.
Edge: Keep camera processing and suitable models on the device. Use remote Ion inference for cloud assistance when the network and latency budget fit the task.

Start with one workflow, its tools, and the channel you want to run.

Open Agent Studio