Agent Release

kit · v1.0

4 entity types · 9 relationships · 6 named queries · 1 guards

Agent release domain overlay composed over the agent-operation base kit. The base supplies the operating layer (Actor, WorkItem, ReviewRequest, Decision, Risk, OpenQuestion, StateNote plus their ownership, review-gate, work-axis, and governed-judgment relationships, queries, guards, and checks). This overlay adds the agent release DOMAIN: agent systems, their versions, eval suites, and receipted eval runs — and one hard gate: an AgentVersion cannot be promoted to live without an accepted, passing, certifying EvalRun.

source on GitHub →

Entity types

AgentSystem

Durable agent, bot, pipeline, or assistant being versioned and released. The stable identity that versions belong to.name · description

AgentVersion

A specific released or releasable state of an agent system: a prompt revision, model swap, tool change, or code build. change_ref points at the change evidence (git sha, config digest, prompt file hash) — the version record does not embed the change itself.label · summary · status · change_ref · created_at

EvalSuite

Named evaluation suite an agent system is judged against. pass_criteria records the human-readable bar (e.g. "grade >= 0.9 on 40 held-out tasks"); the suite definition itself lives with the eval harness, not in state.name · description · pass_criteria

EvalRun

Receipted claim that an eval suite was run against an agent version. Written by the operator or agent after their own harness runs — the kit never executes evals. artifact_ref points at the report/log evidence.ran_at · passed · score · harness · summary · artifact_ref

Relationships

governed9

agent_version_of_system

AgentVersion → AgentSystem

eval_run_scores_version

EvalRun → AgentVersion

eval_run_of_suite

EvalRun → EvalSuite

agent_system_owned_by_actor

AgentSystem → Actor

eval_suite_owned_by_actor

EvalSuite → Actor

eval_run_certifies_version

EvalRun → AgentVersion

agent_version_supersedes_version

AgentVersion → AgentVersion

work_item_targets_agent_version

WorkItem → AgentVersion

risk_attaches_to_agent_system

Risk → AgentSystem

Named queries

certifying_eval_runs_for_version · AgentVersion → EvalRun

Accepted, passing eval runs that certify an agent version. Used by the promotion guard: reads at relationship_state live, so pending certification proposals do not unlock promotion.

agent_version_context · AgentVersion → AnyEntity

From an agent version, inspect its system, scoring and certifying eval runs, targeting work items, and supersession context. all_adjacent expands against the final composed config, so operation seam edges are traversed too.

agent_system_release_context · AgentSystem → AgentVersion

From an agent system, the version line with certification and risk context: per version, counts of scoring runs, certifying runs, and targeting work items.

promotion_candidates · → AgentVersion

Versions awaiting promotion (status candidate) with their certification posture: a candidate with zero certifying runs is not promotable yet.

live_agent_versions · → AgentVersion

Currently live versions with their certification counts.

eval_suite_run_history · EvalSuite → EvalRun

Eval runs recorded against a suite, newest first.

Guards

agent_version_live_requires_certifying_eval · AgentVersion.status -> live

An AgentVersion cannot be promoted to live without at least one accepted, passing EvalRun certifying it (eval_run_certifies_version at relationship_state live). Record the eval run, get the certification accepted through review, then promote.

← all kits Quickstart →