Titan Harness overview
What the Titan Harness is, how it runs your agents on long tasks under Titan's governance, and where your content stays.
The Titan Harness runs your AI agents on long tasks. You define an agent once, give it a goal, and the harness keeps it working in its own sandbox until the task is done: it survives restarts, waits for people when it needs them, stays inside its budget, and reports every step as it happens. Titan decides every action the agent takes.
The harness is its own product, built on top of Titan. It is in private preview: see Private preview access.
What it does
| Feature | What it does | Page |
|---|---|---|
| Agents | A spec bundle (model, instructions, tools, budgets) becomes an immutable version; channels such as prod and canary point at versions | Agents and versions |
| Tasks and runs | A task gives an agent a goal; each attempt and sub-agent is a run; you follow it live, steer it, answer it, pause, resume or cancel it | Tasks and runs |
| Artifacts | Files the agent exports, kept in your organization's storage and downloaded through short-lived links | Artifacts |
| Approvals | Each sandbox tool is a Titan capability: your policy allows it, denies it or asks a person | Approvals and autonomy |
| Budgets and estimates | Spending, step and time limits per task; a price before anything starts | Budgets and estimates |
| Chains | Tasks that start after other tasks and take their outputs | Task chains |
| Fork and replay | Replay a finished run step by step; start a new task from any step of a run | Fork and replay |
| Swarms | Many agents on one goal under one budget: map, lead and workers, pipeline, task board | Swarms |
| Triggers | Tasks on a cron schedule, once at a time, or on demand | Triggers |
| Webhooks | Signed deliveries of the environment's events to your endpoint | Webhooks |
| Attention inbox | The questions and approvals waiting on you | Attention inbox |
| Knowledge and memory | Search your knowledge stores; remember across tasks | Knowledge and memory |
| Data retention | Retention period, zero-retention, legal holds and deletion | Data retention |
| Evidence packs | A zip of a task's Agent Ledger slices, approvals, versions and costs, verifiable offline | Evidence packs |
| CI | Create tasks from a CI job with the job's own identity | Run from CI |
How it fits with Titan
Titan governs; the harness runs. Every action an agent takes through its tools goes through Titan:
- Policy. Each sandbox tool (
shell,read_file,write_fileand the others) is a Titan capability. Your Titan policy allows it, denies it or makes it wait for a person's approval. The harness keeps no approvals of its own. - Identity. Each run acts for the person who created the task. Titan checks that the run's session acts for them before it takes their decision on an approval.
- Record. Titan's Agent Ledger records every decision and execution. An evidence pack exports the part that covers your tasks.
- Controls. Titan's kill switch stops a run's session like any other.
In the private preview, Hexel Studio runs a dedicated Titan instance and a runtime pool for your organization. Your tasks run there, and you run nothing yourself.
Where your content stays
The harness separates what it needs to manage your tasks from what your agents work on.
- The control plane stores metadata only: ids, states, costs, sizes and keyed digests. The text of messages, tool arguments and results, questions and answers stays in your runtime.
- Content is relayed live. While you follow a task, its content reaches your open stream and is not stored on the way. A stream you open later shows the task's history as metadata only.
- Goals and inputs are emptied once a task succeeds; their digests stay. A task that failed or was cancelled keeps its goal and inputs so that it can be retried.
- Files the agent exports go to your organization's storage. A download link points straight at that storage; the bytes do not pass through the harness.
Tasks and runs has the details.
The API
- Base URL:
https://api.hexelstudio.com/v1/harness - Authentication: a Hexel Studio access token for a member or a service user of your organization, as
Authorization: Bearer <token>(see Authenticate). The organization comes from the token. - Scope:
X-Workspace-IdandX-Environment-Idselect the workspace and environment on every request. They never grant access by themselves: your permissions decide (see Permissions). - Writes accept an
Idempotency-Keyheader. Errors are{"error": "<code>", "message": "...", "details": {...}}.
curl https://api.hexelstudio.com/v1/harness/tasks \
-H "Authorization: Bearer $ACCESS_TOKEN" \
-H "X-Workspace-Id: $WORKSPACE_ID" \
-H "X-Environment-Id: $ENVIRONMENT_ID"Every endpoint is in the Titan Harness API reference. You can also use the Hexel CLI (hexel harness ...) and the TypeScript and Python SDKs; the quickstart shows all three.