Limits and spend

Model calls per job. Default 50. A harness agent's session gets it as its turn limit only when it is set; a remote agent's counts the provider's model requests.

Input plus output tokens per job.

How long a job may run, such as 90s, 10m or 1h30m.

The default budget of the agent's jobs, in US dollars, for a job and every job it delegates to. run --budget overrides it for one job. Unset means no budget.

The most each of the agent's model requests may answer, such as 4k: the runtime sends it as the request's max_tokens, and the gateway reserves it against the job's budgets before each call. Unset leaves the model's own default, often its largest output, which a budget then holds in full while each call runs. Pydantic AI agents.

The most the agent's tool definitions may come to in every model request, as an estimate of tokens: every function it lists, or with tool_loading: lazy the two discovery tools, and its built-ins. Unset is 8,000. plugin lint, plugin test, plugin pack and the studio fail past it (the size budget). A whole number from 1 to 1,000,000.

The most the platform adds to the agent's start-up context after its instructions, as an estimate of tokens: the index of its skills, then the start-up text of plugins it uses, such as the memory plugin's accepted decisions and definitions, which gets what the index leaves. Its own instructions do not count. Unset is 4,000. plugin lint, plugin test, plugin pack and the studio fail past it (startup.budget). A whole number from 1 to 1,000,000.

The most memory each sandbox of the agent's sessions may use on a Linux node, swap included, in MiB: the session, and each command it runs in a sandbox of its own (an ACP terminal). Unset is 4096 for a session and 2048 for a command; the tenant's ceiling (oz0 tenant sandbox-limits) is the most it gets. A session that runs out is stopped.

The most processes and threads each of those sandboxes may run; unset is 1024 for a session and 512 for a command. A fork past it fails.

The most CPU time each of those sandboxes gets, in cores (1.5 is one and a half); unset is 2.

The most jobs of the agent that run at once in the tenant, child jobs included, counted for the whole tenant and never per person: a job past it waits at its start until one of them ends, asking again about every 5 seconds, and is never refused for it. It starts at its first ask after a job of the agent completes, fails or is cancelled; a job stopped by a person or timed out frees its place within about 30 seconds.A job that runs under one of the agent's own jobs (its best-of candidates, a quarantine reader, a delegation back to the agent further down) runs in that job's place and takes none of its own, since the job above it waits for it.If the edge cannot count the agent's jobs, a start asks again and goes ahead after five such asks in a row rather than wait for good; an answer that the agent is still at its limit starts that count again, so a start that has waited a long while still waits through one busy answer.A gate's shadow runs do not count. It keeps one busy agent from taking every slot of a shared pool, and works without Temporal's fair dispatch.A whole number from 1 to 1,000; unset is no limit of its own. A flow has none: its steps carry their agents'.

Limits of the whole job tree this agent's job is the root of: model_calls, tokens, children (child jobs in all), fan_out (child jobs running at once), depth and wait (the time each of its jobs may wait for people in all, a duration such as 8h). Zero or unset takes the tenant's (oz0 tenant spend), then the defaults: no limit on calls, tokens and waiting, 500 children, 100 at once, 10 deep. -1 (none for wait) is no limit. See Limits of a whole job tree.

What a loop does (Loops): wait (the default) makes the job wait for a person as an approval of kind loop; stop ends it with LoopDetected; report only reports. Unset, a loop among a harness session's own tools (its shell, reads, edits) is reported, since only repeats in a row count there.

The loop detectors' thresholds: repeats (the same call, 3), cycles (a sequence of two to four calls repeated in a row, 2) and no_progress (results in a row that add nothing, after which only the answer is left, 5).

What a spent job tree's budget does (Budgets that wait): wait for a person to raise it, or fail with BudgetExceeded. Unset takes the tenant's; when neither says, a tree a person started waits and any other fails.

What a write whose outcome is unknown does (Unknown outcomes): a call of a function that is not retry-safe ran out of time, lost its node, or ended with an error the platform cannot read after its request went out.wait asks a person, who answers that it happened, that it runs again under the same idempotency key, or that the job stops; fail fails the job with the failure kind unknown_outcome.The plugin's configuration for the agent overrides this value; when neither says, the tenant's default applies (oz0 tenant spend --on-unknown), and when that is unset too, a job a person started (an account or an operator) waits and any other (a routine's, an event's, an API key's) fails. A flow step's agent keeps its own.

What a model call whose worst case does not fit the tree's budget does (Each call reserves its worst case): last_call (the default) lets it start when nothing else of the budget runs and the budget is not spent, so the tree passes its budget by at most one call's reservation; strict refuses it, so the tree never passes its budget. Unset takes the tenant's (oz0 tenant spend --reserve); last_call does not loosen a tenant's strict.

When the agent is told it is running low and asked to finish: warn_share (0.8), soft_wall (0.85: only the answer is left), and forecast (chance 0.8, min_runs 20, off), which asks a person before the wall from the agent's own finished runs.