Agent harness

Agent harness limits

This page lists the limits and defaults of the agent harness, and what the harness does not do.

Copy page

Agent run defaults

These are the AgentConfig defaults for each run.

Setting	Default
`actions`	`("calendar.create_event",)`
`allow_undo`	`True`
`max_steps`	6
`run_timeout_s`	90 seconds
`spend_cap`	100,000 tokens
`context_token_budget`	4096
`history_turns`	20

RunResult.stopped_reason can be completed, cut_off, max_steps, spend_cap, timeout, model_error or blocked.

Long job limits

Long jobs are off by default. Set long_task=LongTaskConfig(...) in AgentConfig to turn them on.

Setting	Default
Model calls per turn	up to 100
Seconds per turn	up to 3,600
Turns in a chain	up to 10
Spend cap	5,000,000 tokens

On a transient provider error, a long job tries the next model of the preset that the deploy holds a key for.

Memory API limits

These limits apply to the Memory API. See the Memory API errors page for refusals.

Field or limit	Value
`content`	500,000 characters
`label`	200 characters
`id`	512 characters
`metadata`	64 KiB
Batch ingest	1 to 1,000 items
Recall `limit`	1 to 200
Recall `level`	low 8,000, medium 16,000, high 24,000, extra-high 32,000 tokens
Rate limit	Per-minute limit; over it returns 429 with `Retry-After`

Each user trait key is up to 200 characters. Each trait value is up to 2,000 characters.

What the harness does not do

Harny does not host your agent. The agent loop runs on your own model key. The long-job folder and any shell are yours. Harny hosts nothing.
Harny never sees your model key. The model key is read from the provider's usual variable, such as ANTHROPIC_API_KEY, or passed as api_key= with provider=.
Actions are limited to the catalog. The agent can perform only calendar.create_event, email.send, ticket.create, slack.post_message and notion.create_page. Each has an undo except email.send. Undo needs the action record id.
Learning does not write rules. A learned settings offer changes one action kind's setting only when you apply it. You can keep the current setting or restore the previous one. A restore changes the setting for future actions only.
Completing a manual action does not create a rule. It also does not run the step a second time.
Long jobs are off by default. They run only when you set long_task.

Permissions apply to every action. The strictest overlapping setting wins.

Where limits come from

Per-run limits are AgentConfig settings. Memory API limits belong to the Memory API. You set permissions and learning in the console under Actions, Automation.