Skip to content

Errors and limits#

Errors use the format of the Astraeus REST API: {"code", "message", "detail"?}. This page lists Anemoi's codes, then its limits. Errors an agent receives inside its run, from the tool gateway, are in Inside an agent run.

Agents and runs#

Code Status Meaning
INVALID_AGENT 400 The specification is invalid (the message names the field), or a run cannot be made (empty input, the deployment has no endpoint yet).
AGENT_NOT_FOUND 404 No such agent.
AGENT_VERSION_NOT_FOUND 404 The agent has no such version.
AGENT_RUN_NOT_FOUND 404 No such run of this agent.
AGENT_RECEIPT_NOT_FOUND 404 The run has no receipt: it is made when the run ends.
AGENT_ALREADY_EXISTS 409 An agent of that name exists.
AGENT_CONFLICT 409 It was being changed at the same time: try again.
EVAL_GATE_FAILED 409 The version has not passed the gate's suite (or scored below the current version).
BUDGET_EXHAUSTED 409 A budget of the agent or the workspace is spent: no new runs until the period turns over or the budget is raised.
INVALID_POLICY_CHECK 400 A policy check's request is invalid (more than 50 tests…).
INVALID_SANDBOX_COMPOSITION 400 A sandbox composition's request is invalid (an unknown example, a bad base).

A run is also refused as any run is — quotas, an invalid template — with the codes in the Astraeus errors.

Guardrails, approvals, budgets, prices, retention#

Code Status Meaning
INVALID_GUARDRAIL 400 Not Cedar, no forbid, a permit, too long, or too many guardrails.
ORG_GUARDRAIL 403 An organisation guardrail: changed only by the organisation's admins.
GUARDRAIL_NOT_FOUND, GUARDRAIL_ALREADY_EXISTS 404, 409 —
INVALID_DECISION 400 decision is not approve or deny, or the reason is over 500 characters.
NOT_AN_APPROVER 403 You are not among the approval's approvers.
APPROVAL_NOT_PENDING 409 Decided already, expired, or changed meanwhile.
APPROVAL_NOT_FOUND 404 —
INVALID_AGENT_BUDGET 400 The budget's scope, limits or name are invalid.
AGENT_BUDGET_NOT_FOUND, AGENT_BUDGET_ALREADY_EXISTS 404, 409 —
INVALID_RETENTION_POLICY 400 A number of days out of range, or the receipts' floor.

Evaluations#

Code Status Meaning
INVALID_EVAL 400 The suite or the eval run is invalid (the message names the field).
EVAL_SUITE_NOT_FOUND, EVAL_RUN_NOT_FOUND 404 —
EVAL_SUITE_ALREADY_EXISTS 409 —
EVAL_SUITE_IN_USE 409 An agent's gate names the suite: remove it there first.
EVAL_CONFLICT 409 Not possible in the eval run's present state; the message says why.
RUN_ANSWER_UNAVAILABLE 409 A run's answer could not be read from its machine to make a case.

Flows#

Code Status Meaning
INVALID_FLOW 400 The specification is invalid, or an execution's inputs are. POST /flow-checks lists every finding.
FLOW_NOT_FOUND, FLOW_VERSION_NOT_FOUND, FLOW_EXECUTION_NOT_FOUND 404 —
FLOW_ALREADY_EXISTS 409 —
FLOW_CONFLICT 409 Not possible in the execution's or step's present state; the message says why.

Evidence packs#

Code Status Meaning
INVALID_EVIDENCE_PACK 400 Bad name, scope, period (at most 400 days), sections, or the workspace has 100 packs.
EVIDENCE_PACK_NOT_READY 409 Not Ready yet: it cannot be downloaded.
EVIDENCE_PACK_NOT_FOUND, EVIDENCE_PACK_ALREADY_EXISTS 404, 409 —

Marketplace#

Code Status Meaning
INVALID_TEMPLATE 400 Bad slug, metadata, or a specification that looks like it holds a secret.
WORKSPACE_EDITOR_REQUIRED 403 Publishing and installing are for the workspace's editors.
TEMPLATE_MANAGER_REQUIRED 403 Only its managers change, remove or add versions to a template.
ORG_ADMIN_REQUIRED 403 Approving and rejecting are for the organisation's admins.
TEMPLATE_NOT_FOUND, TEMPLATE_VERSION_NOT_FOUND 404 —
TEMPLATE_EXISTS 409 The organisation has a template of that slug: publish a new version of it.
TEMPLATE_VERSION_STATE 409 The version's state does not allow it (the message gives the state).
ALREADY_EXISTS 409 The workspace has an agent or flow of that name: install with update.
NO_CURRENT_VERSION 400 The agent or flow has no current version: name one.

Limits#

Limit Value
Agent name 40 characters
Tools per agent 16; a tool's name 32 characters
Fixed headers per tool 16, each at most 1 024 bytes
Secrets as variables per agent 16
Instructions 64 KiB
Tool policy (with guardrails), sandbox policy 64 KiB each
Model routing rules 8; 32 words each
Input of a run 256 KiB
A run's resources 2 CPU cores, 4 GiB memory, no GPU
A run's time limit 1 hour by default; 60 s to 366 days
Assistant 40 steps; 8 192 output tokens per call; 64 KiB per tool result
Request through the gateway 16 MiB; web answers cut at 5 MiB
Trace arguments kept 2 KiB per call
Approvals waiting per run 16 on its machine
A held call's wait 10 minutes by default, at most 1 day; the approval lapses 24 h after it is asked
A flow's approval step a day by default; 60 s to 7 days
Approval reason 500 characters
Guardrails 64 per workspace, 16 KiB each
Budget overspend window about 20 s of running runs' spend
Traffic split 10 versions; weights sum to 100
Eval suite 200 cases, 10 graders, 512 KiB; 4 cases at a time; 2 h by default (60 s to 24 h)
Flow 64 steps; 3 levels of sub-flows; 10 retries, 20 repeats; outputs 4 KiB per step; events 4 KiB
Flow durations a step 1 s to 30 days; the execution 60 s to 30 days; a sleep 1 s to 7 days
Receipt chain listing 1 000
Evidence pack 400 days, 64 MiB, 100 per workspace; activity of 500 runs
Retention at most 3 660 days; receipts at least 30
MCP server 30 requests a minute per token
Marketplace template 256 KiB of specification