Inside AEC Hub’s governed agentic runtime

AEC Hub now has a governed workflow runtime designed for long-lived project work. The update adds versioned skill, agent, workflow, and evaluation-case definitions, together with durable records for runs, steps, events, artifacts, and professional approvals.

This matters because an agent cannot be governed if its instructions, tools, intermediate state, evidence, and approvals exist only inside a chat session. The runtime moves those controls into inspectable platform objects that are scoped to the firm and retained independently of the model.

The Project Attention Brief reviews sourced project signals such as commitments, decisions, risks, scope changes, and client requests. Deterministic rules surface overdue obligations, open risks, near-term commitments, low-confidence observations, missing evidence, unassigned project signals, and stale connected sources.

The result is stored as a reviewable artifact with retained signal identifiers and evidence counts. A professional can approve it or request changes. Approval completes the AEC Hub run only. It does not send a message, edit a model, update a connected system, or silently promote an observation into project truth.

Each run pins the workflow definition version and stores a snapshot of that definition. Steps declare dependencies, skill versions, inputs, outputs, evidence, and status. The run stops at an approval step that is visible in the same execution tree as the automated work.

Review is recorded atomically against the firm, run, reviewer, decision, and note. Requesting changes closes the current run instead of rewriting its history. A future revision should create a linked run so teams can compare what changed.

The Skills view exposes installed capabilities, their versions, permitted tools, context requirements, review policies, and evaluation fixtures. The Workflows console starts and tracks durable runs. Run detail shows the artifact, evidence coverage, dependency-aware execution tree, lifecycle events, autonomy boundary, and approval gate. Monitoring joins workflow events with source sync and project-signal health.

Proposal drafting is also being routed through this pattern: evidence retrieval can branch in parallel, a grounded draft is versioned, deterministic readiness checks run, and the workflow stops for professional review. The open-ended Assistant remains a query layer and hands repeatable work to governed workflows rather than silently acting.

Reviewer corrections should become versioned evaluation cases so a new skill, prompt, runtime, or model can be compared against the failures professionals actually see. Coverage should state what sources were searched, stale, unavailable, filtered, or outside authorization.

Only after those controls are proven should AEC Hub add model-assisted synthesis and bounded external actions. Those actions should be separate workflows with role checks, exact target previews, idempotency, before-and-after artifacts, provider receipts, and post-action verification.