Skip to main content
agent_config controls how far the agent is willing to go to finish a goal, and how much it’s allowed to spend doing it. Three fields, each independent: mode picks the behavior, max_steps caps tool-call steps, max_duration_seconds caps wall-clock time.

​
Default vs Strict Mode

Default mode is built for open-ended work: browsing, search, and recovery. When the agent hits an unexpected state — a popup, a layout change, a missing element — it explores workarounds before giving up. Strict mode is built for precise, repeatable automation: tests and workflows where an unexpected state should fail the run immediately instead of being worked around. Use it when a workaround succeeding would actually hide a real problem — for example, a checkout test that should fail loudly if a confirmation dialog doesn’t appear exactly where expected, not quietly click past it.
Beta: mode is beta — enroll at agent.tinyfish.ai/beta to use it. Set agent_config.mode to "strict" when creating a run; omit it for the default behavior.

​
Max Steps

max_steps caps how many tool-call steps the agent takes before it stops, independent of mode. Accepts an integer from 1 to 500; runs default to 150 when omitted. Lower it to fail fast and bound cost on simple goals — a one-field extraction shouldn’t need 150 steps, and a low cap catches a goal that’s looping instead of letting it run to the default limit. Raise it for multi-page workflows that legitimately need more steps to finish, such as paginating through a long result set.
Beta: max_steps is beta — enroll at agent.tinyfish.ai/beta to use it. A non-beta request that includes it returns 403 FORBIDDEN; requests that omit it are unaffected either way.

​
Max Duration

max_duration_seconds caps the wall-clock time a run is allowed to take, regardless of how many steps it has used. There’s no limit by default. Unlike mode and max_steps, this field is generally available — no beta enrollment needed. Use it alongside max_steps when a workflow could legitimately need many steps but you still want a hard ceiling on how long you’re willing to wait — max_steps alone doesn’t bound time if each step is slow.

​
Combining All Three

This run fails fast on any unexpected state (strict), stops after at most 30 tool-call steps, and is killed outright if it’s still running after 120 seconds — three independent guardrails, not a single setting.

Configure Agent Mode and Max Steps

Set these from the Playground or the API

TinyAgent Reference

Full agent_config field list and defaults

Runs

What a run’s lifecycle looks like once it starts