Deployment manifest
The deployment manifest AgentBreeder validates and deploys, schema versions 1 and 2: fields, examples and validate. It is not the 2.x agent.yaml format.
Alpha preview. This manifest is not the 2.x agent.yaml (see Legacy 2.x).
It has no framework, prompt, RAG or orchestration fields: your container supplies the agent's
behaviour. To describe an agent instead, write an agent.yaml, which build
turns into this manifest.
Fields
| Field | Contract |
|---|---|
schema_version | Required: 1 or 2. Operator enforce mode requires 2 to deploy |
name | Required lowercase DNS label, 1–63 characters |
image | Required image reference ending in @sha256: plus 64 lowercase hex characters, or a local-only sha256:<64 hex> image ID |
target | local (default) or kubernetes (a local kind cluster only; see Kubernetes on kind) |
port | Integer 1–65535; default 8080 |
health_path | Absolute HTTP path (letters, digits, /, _, -); default /health; no query string |
secrets | Optional map of uppercase environment names to secret://name references; no literal values. Locally, each value is read from a secret file and mounted read-only at /run/secrets/<name> with <ENV>_FILE set |
model_access | Optional: {gateway, aliases, budget_usd?, rpm?}. The operator environment must grant the gateway, every alias and the limits, or the gate denies the deploy (no audit-mode override). With a gateway master key, deploy mints a LiteLLM virtual key for each container generation, limited to the aliases and the fallbacks the environment grants for them, and revokes it on remove. agentbreeder build writes this block |
mcp_access | Optional, needs model_access: [{ref: mcp/<name>@X.Y.Z, tools: [...]}]. The environment must grant every tool to the agent (MCP_TOOL_NOT_GRANTED). Deploy mints a relay token per container generation. Refused with runtime_enforcement: mcp_http (MCP_RELAY_UNSUPPORTED); see MCP tools |
guardrails | Optional, needs model_access: up to 10 agent guardrail names the environment offers (GUARDRAIL_UNDEFINED otherwise). Deploy attaches them to the deployment's gateway key; see Guardrails |
identity | v2, required: provider (operator connection name), ref (provider-specific immutable ID), environment, optional expect_owner. A reference only; AgentBreeder never creates identities |
requires.roles | v2: list of {authority, resource, role_id} immutable IDs that must already be granted |
requires.groups | v2: list of {authority, group_id} stable IDs |
requires.attributes | v2: map of namespace/name to an exact required value |
capabilities | v2: tools: [{mcp: https://…, actions: [...]}] and approval_required. A deploy-time ceiling; enforced per call only when the operator turns on runtime enforcement |
policy.restrictions | v2: operator-approved restriction bundle IDs. They can only restrict |
Input must be a single UTF-8 YAML mapping of at most 64 KiB. Unknown fields, duplicate keys, aliases/anchors and wrong types are rejected.
Nothing in the manifest can grant a permission, select the baseline policy, choose the enforcement mode or create an environment. Those belong to the operator configuration.
Examples
schema_version: 1
name: support-agent
image: registry.example.com/support@sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
target: local
port: 8080
health_path: /healthschema_version: 2
name: refund-agent
image: registry.example.com/refund@sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
target: local
identity:
provider: corp
ref: 4f1c2b7e-0000-4000-8000-000000000001
environment: prod
expect_owner: alice@example.com
requires:
roles:
- {authority: tenant-id, resource: payments-api-id, role_id: refunds-role-id}
groups:
- {authority: tenant-id, group_id: finance-group-id}
attributes:
acme-governance/data_classification: confidential
policy:
restrictions: [pci]Validate
.venv/bin/agentbreeder validate agent.yaml # exit 0 valid, 2 invalid
.venv/bin/agentbreeder validate agent.yaml --json # normalized outputValidation doesn't deploy anything or check that the image or secrets exist.
Canonical reference: docs/manifest.md and its JSON Schema.
Identity gate
agentbreeder deploy refuses an agent that is not onboarded in your identity provider. Policy, approvals, a kill switch, the audit log and the MCP egress sidecar.
Legacy 2.x
AgentBreeder 2.x is what a bare pip install gives you today. It is unmaintained, superseded by the focused rebuild, and has known functional and security issues.