HarnessGet started
What Harness is
Harness is where you build an Ailo. Who it is, which procedures it follows, what it can look up and which tools it can call are all written in Harness and saved as immutable versions; testing, going live, rolling back and reading run records happen here too. The conversation itself lives in Desk: Desk decides who is serving, what the customer sees and which actions are allowed.
This page covers the concepts in Harness and the layers a reply passes through, from the customer's message to the message going out. To get going, see Build your first Ailo.
Concepts #
| Concept | What it is |
|---|---|
| Organization | Your account space in Harness. Desk and Harness each have their own organizations, with separate data |
| Ailo | An agent that can serve customers. Each Ailo is one package |
| Package version | Everything an Ailo is at one point in time: system prompt, SOPs, FAQ, tools, engine parameters. Never changed once saved |
| Live version | The package version serving customers now. Only going live and rolling back change it |
| Turn | Ailo's handling of one trigger, usually a new customer message. It pins the live version when it starts |
| Code package | The code a turn actually runs, entered through main(context, handle). See Code packages and main |
The layers a reply passes through #
Text
The customer writes
→ Desk opens a turn and hands the Runtime this turn's context and permissions
→ The Runtime pins the live package version and calls the code package's main(context, handle)
→ main uses the SDK to call the model and tools, and hands what it wants to say back to Desk
→ Desk checks who is serving, the evidence and the site's rules, then sends, sends for review or refuses
| Layer | Responsible for | Cannot |
|---|---|---|
| Desk | Conversation facts, who is serving, running business tools for real, checking and reviewing customer replies | Run Ailo's model loop or write Ailo's replies |
| Runtime | Calling main, calling the model, the call and budget limits of each turn, run records |
Decide for Desk what the customer sees |
| Code package | Choosing instructions and content, building the model input, running its own tools, handling Desk's receipts | Fake tool results or progress, hold Desk's credentials, or get around Desk's handoff rules |
| SDK | A ready-made model loop, tool definitions, requests to Desk | Use any private channel around the Runtime or Desk |
What the platform always guarantees #
- The model's raw text never goes straight to a customer. Anything said to a customer goes to Desk through the reply tool and is sent only if Desk accepts it.
- Desk is the last gate. Serving rights, evidence checks and review rules apply to every code package alike; a package's own approval counts for nothing.
- Times, amounts and outcomes need evidence. Evidence comes from data or knowledge entries actually retrieved in this turn, and Desk checks it.
- Internal content stays internal. Handoff notes for the team, ticket details, model reasoning and raw tool arguments never reach the customer.
- The site's configuration governs the handoff to a person. A package version can only narrow it (turn it off or limit the mode); it cannot turn it on where the site has it off, and a site override can force it on. Code in a package cannot get around Desk.
What is available now #
| Capability | Status |
|---|---|
| Built-in loop | The live default, used when a package version's engine.main is unset or loop |
Standard package @ailodesk/harness-standard |
Optional: set engine.main to standard. Runs through a temporary adapter for now and is off by default |
| Custom code packages | Cannot run in live conversations yet. You can upload a ZIP, have it statically checked and save it as a version. See Code packages and main |
SDK @ailodesk/ailo-sdk |
Lives in the ailodesk repository; the standard package is written with it. Not yet published as an npm package you can install on its own |
| Project template | Planned, not available yet |
| Content: system prompt, SOPs, FAQ | Read, edit and publish through the API. See Knowledge and SOP API |
| Test bench, release, runs, stats, Orbit | Available. See The Harness console |
| Evals page | Not available yet; the page is a placeholder |