harnyAgent harness
Agent harness
The agent harness is what lets an agent work for your users inside your app. It remembers each user, reaches the accounts each user connects, acts only inside the permissions you set, hands steps that need a person to a person, and records every run. You can also use Harny for memory only: the Memory API works without the harness.
What the agent harness is
Your agent talks to your users inside your app. The harness gives it what it needs to work for them: memory for each user, the accounts each user connects, permissions you set, and a record of every run.
The limits sit in your console, not in your agent's prompt. The harness does not choose your agent's model. Your model runs the agent, on your key.
What your agent gets
| Part | What it does | Where you control it |
|---|---|---|
| Memory | Keeps what your agent knows about each user, per user. | Memories |
| Connected accounts | Lets your agent reach the accounts each user connects. | Connections |
| Permissions | For each of four kinds of action, you choose Act without asking, Ask the user first or Never allowed. | Actions › Automation › Permissions |
| Manual actions | Holds steps your agent routed to a person, such as a signature, a call or access. You record each as done. | Actions › Manual actions |
| Learning | Offers suggestions drawn from recorded responses. You apply them or keep your settings. It is not model training. | Actions › Automation › Learning |
| Instructions | Stores a saved request to your coding assistant. It is not a rule that runs by itself. | Actions › Automation › Instructions |
| Activity | Records every run: what it was asked, what it did, how long it took. | Actions › Activity |
| Long jobs | Lets one agent work through a folder of files for up to an hour. It is off by default. | Actions › Automation › Permissions, one switch |
Getting there
- Create an app key in the console.
- Choose how to connect: paste the one-line setup prompt into your coding agent, or connect over MCP at
https://api.harny.ai/v1/mcp/mcpwith your app key. - Set a mode for each of the four kinds in Actions › Automation › Permissions: "Nobody else affected", "Others affected", "Financial operation" and "Irreversible".
- Run a first job. Ask your agent to move the Acme pilot budget for Nadia from 32k to 40k, then watch it work.
- Open Actions › Activity to see what the run was asked and what it did.
Frequently asked questions
01Whose model runs the agent?
Yours. Harny does not choose your agent's model. Your model, your key.
02What may the agent do without asking?
Only the kinds you set to "Act without asking". Kinds set to "Ask the user first" wait for the user. Kinds set to "Never allowed" do not run.
03When is a person needed?
When a step needs a signature, a call or access, your agent routes it to a person. It appears in Manual actions, and you record it as done.
04Does it cost extra beyond users?
Harny meters monthly-active end users, not agent runs. A person counts in a month when they used your app that month. The plans are on the pricing page.
05Where is the record?
In Actions › Activity. Every run shows what it was asked, what it did and how long it took.
06Can I use only memory without the harness?
Yes. The Memory API works on its own. The harness parts switch on as you use them.