AIWG 2026.8.17 - Runnable graphs, useful release conversations

Release date: 2026-08-22 Channel: stable

AIWG 2026.8.17 makes graph-shaped agent work concrete without turning graphs into the answer to every orchestration problem. The composition runtime can now execute validated Flow graphs with typed state, bounded joins, retries, replay, and capability narrowing. The important part is not that a diagram exists. The important part is that the same graph can be validated, run, inspected, and replayed with evidence about what actually happened.

This release also changes how we talk about releases. The changelog remains the durable record, but stable releases now get a GitHub Announcements discussion that focuses on what the changes mean in practice and gives people a place to compare how they are using them.

What this changes for you

You can run graph compositions, not just describe them

`aiwg composition run` executes phased and multi-track `FlowGraph` definitions through an explicit adapter. Activation and invocation IDs are stable, joins are bounded, reducers are typed, retries are mutation-safe, and checkpoints can be replayed. Requested resources and realized resources are recorded separately so the trace shows what the runtime actually used.

This helps when dependencies, parallel branches, joins, or recovery paths need to be explicit. A loop is still usually the better tool for one focused task. Think of the graph as the road map between work units; the agents and loops are still the vehicles doing the work.

Start by validating a graph before running it:

aiwg composition validate path/to/flow-graph.yaml
aiwg composition run path/to/flow-graph.yaml

Composition policies can be compared with receipts

The new benchmark expands one fixed-task manifest across single-pass, Self-Refine, parallel-candidate, strict LCM, adaptive-convergence, and budget-partial policies. It reports success-conditioned resource use, injects known failure modes, and blocks quality or efficiency claims when the evidence does not support them.

If you are deciding whether another agent pass or another parallel branch is worth the cost, this gives you a repeatable comparison instead of a preference disguised as a benchmark.

The graph pattern now has a practical development map

The new gap audit separates what AIWG already has from what graph-oriented development still needs. AIWG already owns the Flow DAG, gates, fanout, capability routing, Mission durability, A2A dispatch, and audit substrate. The next work is narrower: conditional routes, guarded cycles, reducers, shared run identity, runtime adapters, graph conformance, and an operator view in Cockpit.

That distinction matters. Building a second graph DSL would create two systems that slowly disagree. Extending Flow keeps one execution and evidence model.

Release notes now have a conversation layer

For stable releases, publication verification now completes before the release flow opens a GitHub Announcements discussion. That discussion links back to the GitHub release, all published npm versions, these notes, and the exact CHANGELOG section. It highlights practical impact and asks concrete questions rather than copying every bullet.

The goal is simple: release records tell you what shipped; the discussion helps work out how to use it well.

Cockpit browser bootstrap is easier to recover

If a Cockpit Bridge is already running and its printed one-time browser URL has expired, the installation guide now provides a command to mint a fresh URL. It also explains source-checkout and SSH local-forward use without putting the reusable Bridge bearer into browser history or forwarded URLs.

Other improvements

• Static Flow site deployment now has a validated private-origin plan with

explicit DNS, Caddy, volume, workflow, and operator handoff boundaries.

• Built-in codebase indexing recognizes conventional Python manifests, package

roots, tests, and scripts while preserving JavaScript and TypeScript defaults.

• Provider transformation receipts now record unavailable evidence explicitly

instead of implying evidence was collected when the provider could not expose it.

• Cockpit deployment diagnostics, session replay, Codex dangerous-launch

handling, attestation transport conformance, and mutation-test native reloads received additional hardening.

Package matrix

[email protected]
@aiwg/[email protected]
@aiwg/[email protected]

Upgrade

Upgrade the package set together, then regenerate provider context:

npm install --global [email protected]
npm install --global @aiwg/[email protected]
npm install --global @aiwg/[email protected]
aiwg regenerate

Rollback

Roll all three npm packages back together to `2026.8.16`, then regenerate provider context. Preserve graph-run checkpoints and release-verification evidence so the rollback does not erase the record of work already executed.