Skip to main content
Engineering intelligence is the part of ChangeGuard AI that reads like a senior engineer: it adds context to a pre-flight judgment, takes an owned position on an incident, shows the reasoning behind its confidence, and recalls what happened the last time something similar broke. It is generally available since the release of 7 November 2026. It never decides for you. A verdict stays SHIP, HOLD or BLOCK on the evidence; advice stays advice; execution still needs a Permit.

The four capabilities

The two collaboration capabilities released with them are documented under Investigate: the incident workspace and the Code Graph.

Grounded in evidence, explicit about uncertainty

Every claim an opinion or an advisor note makes is bound to a record ChangeGuard AI holds: an incident, a change record, a root-cause analysis, an evidence line, a remediation outcome. A reason that cannot be attributed to evidence is dropped before you see it; an opinion with no attributable reason is rejected and not stored. Uncertainty is part of the output, not a footnote:
  • The Advisor says what it cannot see (blindspots) and stays silent below its materiality floors.
  • The Opinion carries a confidence between 5 and 95 percent, stated in words, with the open unknowns and the falsifier (what would change the call).
  • Reasoning factors are derived by the backend from the signals it has; they cannot weigh evidence the platform never saw.
  • Experience states similarity as similarity, never identity, and always with the backend’s exact counts.
Where there is no evidence to stand on, the capability stays quiet: an Advisor with no live environment evidence reports no_environment_evidence; an Opinion forms nothing without a root-cause analysis; Experience recalls nothing from a workspace without recorded outcomes.

Who gets it, and how to control it

Availability is decided per tenant and per request from three inputs: the plan, your own setting, and the platform switch. Each capability is on by default where the plan carries it; there is nothing to request or enrol in.

When the model is unavailable

Judgments, the CSC Score, Permits, execution and verification make no model call and are unaffected by any model’s availability. So is the Engineering Advisor: it is a projection over records the platform already holds (active incidents and the recent change record) and makes no model call. The Opinion, with its Reasoning and Experience, reaches Amazon Bedrock through one router with a preferred model and a permanent, evaluated fallback; when neither can answer, the product says so in plain words and nothing is stored. See Model availability.

Challenge it

You are supposed to push back. Every opinion lists the evidence it rests on; the incident conversation records what was observed and decided, in order; a re-run of the pre-flight updates a verdict when reality changed; and an override you make is recorded as evidence against the judgment’s track record. See How ChangeGuard AI makes decisions.