Every tool means the same thing by “run”
Workflows, runs, steps, tool calls, evaluations and incidents get one definition, so producers and consumers stop writing adapters for each other.
v0.1 – v0.3AI Operations Specification · Roadmap
The road to a stable standard for AI operations.
Eight milestones, built one layer at a time and reviewed in the open, ending in a spec that anyone can implement without our packages.
Step 1 of 8 · v0.1 in review · v0.2–v0.4 drafted · updated Oct 2026
One contract for how AI systems run, cost and fail.
Every milestone moves the spec closer to four promises.
Agent runs as graphs, not flat logs
Retries, branches, parallel steps and handoffs are explicit edges. Timestamps alone are never allowed to prove causation.
v0.2Artifacts you can validate yourself
JSON Schemas plus semantic checks, with a conformance suite you can run without installing a DeepAgentLabs package.
v0.4 · v0.6.xBuilt to outlive any single vendor
Independent review gates every release, and clear versioning and extension rules mean anyone can build on it safely.
v0.5 – v1.0The milestones
Eight steps from a draft to a stable standard.
v0.1 · Core concepts In review
Define what exists inside a running AI system.
The foundation: a shared meaning for every runtime occurrence, with clear boundaries between them.
- A Workflow is a definition; a Run is one execution of it.
- Model calls, tool calls, retrieval and memory stay distinct occurrences.
- A reliability event is evidence. It doesn't automatically become an Incident.
Done when two independent teams classify the same runtime events the same way.
WorkflowRunRequestStepAgentModel interactionPromptTool invocationRAG retrievalMemory operationEvaluationSafety signalReliability eventIncident
14 concepts · internal decisions resolved · awaiting independent review
v0.2 · Relationships Exploratory draft
Show how the pieces connect into execution graphs.
Real agent runs branch, retry, loop and hand off. v0.2 makes every one of those relationships explicit.
- Workflow → Run → Request hierarchy and parent–child Runs.
- Sequential, parallel, branch/join, retry and loop semantics.
- Agent participation and handoff patterns.
Done when two tools draw the same execution graph.
Explicit edges: containment, ordering, causation, dependency, handoff
v0.3 · Semantics Exploratory draft
Give every event one name and one meaning.
Canonical, transport-neutral events with clear lifecycles across runs, agents, models, tools, memory, evaluation, safety and incidents.
- Canonical event names and naming rules.
- Lifecycle meanings: what started, finished or failed.
- The minimum attributes needed to interpret each event.
Done when producers and consumers agree on what each event means.
- run started
- step started · plan
- model interaction completed
- tool invocation started
- tool invocation failed · timeout
- reliability event recorded
- evaluation recorded · pass
Same meaning on any transport: OTLP, files or queues
v0.4 · JSON Schemas Exploratory draft
Turn the spec into artifacts that machines can check.
Workflow and Run schemas, plus validation fixtures, all derived from v0.1–v0.3. Two conformance artifacts already pass.
- Structural validation: is the artifact the right shape?
- Semantic validation: do references resolve, and is every graph acyclic?
- Passing today: AgenticLens and Agentic Chaos sample runs.
Done when artifacts validate without any DeepAgentLabs package.
{
"artifact_type": "run",
"steps": [ … ],
"relationships": [ … ],
"events": [ … ]
}
✓ Schema valid✓ References resolve✓ Graph acyclic
v0.5 · Versioning Planned
Know whether an artifact is safe to consume.
Version markers and change rules so the spec can evolve without breaking the tools that read it.
- Artifact version markers.
- Additive vs. breaking change rules, and a deprecation policy.
- Expected behaviour when a consumer meets a newer version.
Done when anyone can tell if an artifact is safe to consume.
- New optional fieldadditive · safe
- Field marked deprecatedwarn · still works
- Required field renamedbreaking · new major
- Unknown extension namespacepreserve · safe
v0.6 · Examples & extensions Planned
Let anyone extend the spec without forking it.
Canonical example artifacts, from minimal to advanced, plus clear rules for third-party extensions.
- Minimal and advanced canonical artifacts.
- Reverse-domain namespaces for vendor data.
- The spec owns conformance; tools only make it convenient.
Done when third parties can extend the spec without forking it.
"attributes": { "io.deepagentlabs.agenticlens.cost_usd": 0.0042, "com.yourcompany.team": "support" }
Unknown namespaces are preserved, never rejected
v0.6.x · Provenance & evidence Planned
Show where every finding came from.
Attach source references and lineage to findings, and give producers a conformance suite they can run on their own.
- Provenance concepts: sources, lineage, derived findings.
- An independent conformance test suite and fixtures.
- Next: OpenTelemetry GenAI binding, audit and verification schemas, hosted docs.
Done when producers can run conformance checks on their own.
- FindingRefund agent skipped policy check
- Evaluationpolicy-adherence · fail
- Steptool call · refunds.create
- Sourcetrace 7f3a… · AgenticLens
v1.0 · Stable standard North star
Freeze it so anyone can build on it.
The reviewed core model, relationships, conventions, schemas, versioning, examples and extension model, all locked into one stable release.
Done when third parties implement the standard without importing our packages.
AIOS 1.0
Stable · vendor-neutral
What unlocks the next step.
01
Independent review
People outside DeepAgentLabs review v0.1–v0.4. The invite packet is ready.
02
Outside implementation
A producer or consumer we didn't build proves the v0.4 format works.
03
Compatibility policy
Written change rules come before anything is called stable.
Shape what comes next
Review the spec before it's frozen.
v0.1 is waiting on independent reviewers. If you build agent frameworks, evals or reliability tooling, your review moves the roadmap forward.