Limits

The numeric bounds on task execution, governance and the authorization graph. Every value below is the value in source; overrides are listed per limit.

Synopsis

A limit comes from one of three places, and that decides how you change it.

Source	Changed by	Scope
Policy rule	`POST /v1/policies`, or `task_policy` at task creation	Per workspace, agent, user or task, at runtime
Environment setting	Process environment or Helm values	Per deployment, at restart
Code constant	Editing the source and redeploying	Global

Fields

Category	Source	Notes
Policy ceilings	Policy rule	Merge monotonically across four layers; a lower layer may only tighten.
Workflow and worker settings	Environment, prefix `WORKFLOW__`	`EXECUTION_ENGINE=temporal` makes the three `TEMPORAL_*` connection settings required with no defaults.
Access-control settings	Environment, prefix `ACCESS_CONTROL_`	See Defaults and overrides.
Agent-loop limits	Persisted effective policy	Required at task admission; runtime code has no numeric fallback.
Timeouts and retries	Environment or Temporal activity configuration	Per deployment.
Context and output sizes	Code constant	Not configurable.
Event-log bounds	Code constant	Apply to event projections only.
Pagination caps	Code constant	Clamped server-side.

Values

Policy ceilings

Limit	Default for a new workspace	Unit
`budget.monthly_spend_cap_usd`	`500.00`	USD per calendar month, UTC
`budget.run_budget_usd`	`50.00`	USD per task
`budget.service_budget_usd`	not seeded	USD per task
`tokens.max_tokens`	`20000000`	tokens per task
`tokens.max_tokens_per_call`	`100000`	maximum output tokens for one LLM call
`execution.max_model_turns`	`100`	model turns per task
`execution.max_tool_calls_per_turn`	`10`	tool calls per model turn
`execution.max_tool_calls_total`	`1000`	tool calls per task

Execution limits

The effective policy must contain run budget, total and per-call token ceilings, model turns, per-turn tool calls, and total tool calls. Missing dimensions reject task execution rather than selecting a code default. The values above are seed data for a new workspace and become rows in the policy database; operators may replace the seed file and existing workspace rules are not silently overwritten.

Guard thresholds

Warning fires at the ratio; denial fires at full consumption.

Guard	Warning threshold	Phase
`CostBudgetGuard`	0.8	pre-LLM and pre-tool
`ServiceBudgetGuard`	0.8	pre-tool
`TokenBudgetGuard`	0.85	pre-LLM

Timeouts

Limit	Value
`ACTIVITY_TIMEOUT`	5 minutes
`LLM_CALL_TIMEOUT`	10 minutes
`TOOL_EXECUTION_TIMEOUT`	35 minutes
`DELEGATION_TIMEOUT`	10 minutes
`EVENT_PUBLISH_TIMEOUT`	5 seconds
`HEARTBEAT_TIMEOUT`	30 seconds
`CONTINUATION_TIMEOUT`	24 hours
`WORKFLOW__AGENT_VALIDATION_TIMEOUT_MINUTES`	5
`WORKFLOW__AGENT_EXECUTION_TIMEOUT_HOURS`	24
`WORKFLOW__DYNAMIC_ACTIVITY_TIMEOUT_MINUTES`	30
`WORKFLOW__TEMPORAL_MAX_WORKFLOW_DURATION_DAYS`	7
`ACCESS_CONTROL_OPENFGA_TIMEOUT_SECONDS`	10.0

There is no timeout on a pending human approval.

Retries

Limit	Value
`DEFAULT_RETRY_ATTEMPTS`	3
`LLM_RETRY_ATTEMPTS`	3
`EVENT_PUBLISH_RETRY_ATTEMPTS`	1

LLM calls retry transient failures with backoff. Permanent failures — auth, quota, billing, an unknown model — fail fast via a non-retryable flag.

Concurrency

Limit	Value
`WORKFLOW__TEMPORAL_MAX_CONCURRENT_ACTIVITIES`	10
`WORKFLOW__TEMPORAL_MAX_CONCURRENT_WORKFLOWS`	5

There is no per-workspace request rate limit, task concurrency quota, or tool call rate limit.

Context window management

Limit	Value
`DEFAULT_CONTEXT_WINDOW`	128000 tokens, when the model does not declare one
`CONTEXT_COMPACT_THRESHOLD`	0.75
`CONTEXT_WARNING_THRESHOLD`	0.60
`CONTEXT_RESERVE_FOR_OUTPUT`	0.15
`MIN_RECENT_MESSAGES_TO_KEEP`	6
`TOKENS_PER_MESSAGE_OVERHEAD`	4 tokens

Output offloading

Limit	Value
`TOOL_OUTPUT_OFFLOAD_CHARS`	8000 characters
`OUTPUT_SUMMARY_HEAD_CHARS`	500
`OUTPUT_SUMMARY_TAIL_CHARS`	200
`READ_OUTPUT_MAX_RETURN_CHARS`	16000
`HISTORY_SEARCH_MAX_RESULTS`	20

Event-log bounds

Applied to Redis and database event projections. Activities and the model still receive the original values.

Limit	Value
Maximum string length kept inline	2000 characters
Base64-like string omitted at	256 characters
Control-character ratio treated as binary	above 0.02
Maximum nesting depth	10
Maximum dictionary entries kept	100
Maximum list items kept	100

Field names matching a secret pattern are replaced with [redacted]. Fields carrying content, command text, stdout or stderr are replaced with a size marker.

Pagination

Endpoint area	Default	Maximum
Audit logs	50	100, validated and clamped again server-side
OpenFGA bootstrap store and model listing	100 per page	—

Enforcement

Limit	Enforced at
`monthly_spend_cap_usd`	Task creation only. A running task is not interrupted, and the check has no lock, so concurrent creation can cross it.
`run_budget_usd`	The workflow loop, and `CostBudgetGuard` before each LLM and tool call.
`service_budget_usd`	The workflow loop, and `ServiceBudgetGuard` before each tool call.
`max_tokens`	`TokenBudgetGuard` before each LLM call.
`max_tokens_per_call`	LLM call resolution; the strictest of policy, request, and model capability wins.
`execution.max_model_turns`, `execution.max_tool_calls_per_turn`, `execution.max_tool_calls_total`	The workflow loop and tool-call budget.
Timeouts and retries	Temporal activity options.
Event-log bounds	Event projection, before publish and persist.

Condition	Behaviour
A budget ceiling absent or not greater than zero	The guard allows unconditionally. `null` is not zero.
A required runtime policy dimension is missing	Task execution fails closed with the missing field named.
A service budget of zero or less	Treated as unlimited.
A guard raising an exception	Logged with a traceback; the pipeline continues. Fail-open for that dimension.
Month-to-date spend	Summed from persisted task cost from the first of the current UTC month. Spend from a task still running is not fully counted.
Per-call cost	Read from the provider response where supplied; otherwise estimated at a flat rate per token.

The max_depth parameter on the graph client’s check accepts a value and defaults to 10, but it is not sent in the request. Setting it has no effect.

Defaults and overrides

Where a default exists in more than one place, the code default is authoritative for a process started without the corresponding environment variable.

Setting	Code default	`docker-compose.dev.yaml`	`docker-compose.yaml`	Helm
`ACCESS_CONTROL_BACKEND`	`disabled`	`openfga`	absent — code default applies	`openfga` when `openfga.enabled=true`, the chart default
`ACCESS_CONTROL_OPENFGA_AUTO_BOOTSTRAP`	`false`	`true`	absent	`true`
`ACCESS_CONTROL_OPENFGA_AUTO_APPLY_MODEL`	`false`	`true`	absent	`true`
`ACCESS_CONTROL_OPENFGA_STORE_NAME`	`agentarea`	`agentarea`	absent	—
`ACCESS_CONTROL_OPENFGA_API_URL`	`http://openfga:8080`	`http://openfga:8080`	absent	—

make up runs docker-compose.yaml and make up-dev runs docker-compose.dev.yaml. On the make up path the graph backend is disabled, in which case every permission check returns allow.

Setting	Code default	Override
`GOVERNANCE_DEFAULT_POLICIES_PATH`	unset, uses the packaged `config/default_policies.yaml`	Path to an alternative defaults file. A missing file means new workspaces start with no rules.
`WORKFLOW__EXECUTION_ENGINE`	`temporal`	`direct` ignores the `TEMPORAL_*` settings.
`WORKFLOW__TEMPORAL_SERVER_URL`	`""`	Required when the engine is `temporal`; startup validation fails if missing.
`WORKFLOW__TEMPORAL_NAMESPACE`	`""`	Required when the engine is `temporal`.
`WORKFLOW__TEMPORAL_TASK_QUEUE`	`""`	Required when the engine is `temporal`.

Every constant listed under Values that is not shown with an environment prefix is a code constant with no override.

Example

Tighten one task below the workspace ceiling. The effective run budget becomes the minimum of the request value and every layer above it:

{
  "description": "Summarise the Q3 report",
  "task_policy": {"budget": {"run_budget_usd": "2.50"}}
}

Start here

Concepts

Guides

Self-host

Reference

API Endpoints

Community

Limits

Limits

Synopsis

Fields

Values

Policy ceilings

Execution limits

Guard thresholds

Timeouts

Retries

Concurrency

Context window management

Output offloading

Event-log bounds

Enforcement

Defaults and overrides

Example

See also

​Limits

​Synopsis

​Fields

​Values

​Policy ceilings

​Execution limits

​Guard thresholds

​Timeouts

​Retries

​Concurrency

​Context window management

​Output offloading

​Event-log bounds

​Pagination

​Enforcement

​Defaults and overrides

​Example

​See also

Limits

Synopsis

Fields

Values

Policy ceilings

Execution limits

Guard thresholds

Timeouts

Retries

Concurrency

Context window management

Output offloading

Event-log bounds

Pagination

Enforcement

Defaults and overrides

Example

See also