v2026.8.17

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

aiwg@2026.8.17
@aiwg/cli@2026.8.17
@aiwg/cockpit@2026.8.17

Upgrade

Upgrade the package set together, then regenerate provider context:

npm install --global aiwg@2026.8.17
npm install --global @aiwg/cli@2026.8.17
npm install --global @aiwg/cockpit@2026.8.17
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.