goal | string | yes | | length 1 to 4096 | What the agent should achieve, in your words. |
instructions | string | | | up to 16384 characters | How to go about it: constraints, preferences, what to report. |
endpointRef | reference to InferenceEndpoint | yes | | | The chat InferenceEndpoint (the model) the agent thinks with. |
roleRef | object | yes | | | The Role the agent holds in this project: its tools. It may not grant access control, billing, secret values, status or wildcard writes. Setting it needs the right to bind this role. |
roleRef.name | string | yes | | length 1 to 63 | Name of a Role in this project or a platform role. |
state | string (one of) | | "Running" | Running, Paused, Stopped | Running works; Paused keeps the agent's place and calls nothing; Stopped ends it at once (the kill switch). |
extraClusters | list of object | | | up to 6 items | Other clusters the agent may read (never write). Only for agents of the platform's own agents project, and only the cluster platform and the projects of its swarm's change repositories. |
extraClusters[].name | string | yes | | length 1 to 63 | The cluster (only platform). |
extraClusters[].roleRef | object | yes | | | A Role that grants read verbs only; the agent holds it in that cluster. |
extraClusters[].roleRef.name | string | yes | | length 1 to 63 | Name of the Role. |
triggers | object | | {} | | When the agent works. |
triggers.mode | string (one of) | | "AlwaysOn" | AlwaysOn, Schedule, Events | AlwaysOn loops and waits intervalSeconds when it has nothing to do; Schedule works at the cron times; Events works when a watched object changes. |
triggers.intervalSeconds | integer | | 300 | 30 to 86400 | How long the agent waits after it has nothing more to do. |
triggers.cron | list of string | | | up to 8 items | Cron expressions in UTC, for Schedule. |
triggers.watches | list of object | | | up to 16 items | Objects of this project whose changes wake the agent, for Events. |
triggers.watches[].group | string | yes | | up to 253 characters | API group of the kind. |
triggers.watches[].resource | string | yes | | up to 63 characters | Plural resource name, for example vpcs. |
triggers.watches[].name | string | | | up to 63 characters | One object; all of the kind when empty. |
triggers.watches[].cluster | string | | | up to 63 characters | The cluster of the objects; the agent's own project when empty (otherwise one of its extraClusters). |
budget | object | | {} | | Limits the agent stays within. |
budget.tokensPerDay | integer | | 200000 | 1 to 10000000 | Model tokens (input and output) per UTC day. |
budget.cpuSecondsPerDay | integer | | 600 | 1 to 86400 | CPU seconds per UTC day spent on the agent's behalf. |
budget.actionsPerHour | integer | | 20 | 1 to 600 | API calls per hour. |
budget.stepsPerRun | integer | | 50 | 1 to 1000 | Steps before the agent rests, as if it had nothing more to do. |
budget.spendCapMicros | integer | | 0 | at least 0 | Money the agent may spend, in millionths of the account currency. 0 unless a billing owner sets it. |
memoryRefs | list of Memory | | | up to 8 items | Memories the agent may recall from and write notes to (its role must also allow memories/records). |
memoryRefs[].name | string | yes | | length 1 to 63 | Name of the Memory in this project. |
approval | object | | {} | | Which actions need a person's approval before they run. |
approval.requiredFrom | string (one of) | | "high" | low, medium, high, never | Actions of this risk and above need approval: low is reading, medium is creating and changing, high is deleting. never asks for nothing. |
approval.alwaysAsk | list of object | | | up to 32 items | Kinds whose writes always need approval. |
approval.alwaysAsk[].group | string | yes | | up to 253 characters | API group of the kind. |
approval.alwaysAsk[].resource | string | yes | | up to 63 characters | Plural resource name. |
approval.expiresAfterSeconds | integer | | 86400 | 60 to 2592000 | A request nobody decided on expires after this long. |
monitoringRef | reference to MonitoringWorkspace | | | | Optional MonitoringWorkspace for the agent's step metrics. |