Building ChangeMesh: Rehearsing Enterprise Changes Before They Become Incidents

#ai #googlecloud #devops #hackathon

This article was created specifically for the purpose of entering ChangeMesh in the All Things Agentic Hackathon.

Building ChangeMesh: Rehearsing Enterprise Changes Before They Become Incidents

Enterprise change is rarely a single edit.

A database schema rename, API contract change, or migration can cross repositories, services, policies, ownership boundaries, and deployment systems. A coding agent may be able to modify files, but that does not prove the overall change is safe.

That is why I built ChangeMesh.

ChangeMesh is a proof-carrying multi-agent system that treats enterprise change as a long-lived, recoverable distributed transaction rather than a chat session.

Its core loop is:

Discover → Qualify → Rehearse → Ground → Authorize → Execute → Prove → Certify

A six-agent Google ADK fleet

ChangeMesh uses a six-agent Google Agent Development Kit (ADK) architecture.

Gemini 3.6 Flash, running through Vertex AI and the Google GenAI SDK, performs bounded semantic reasoning.

But model judgment is deliberately not allowed to become execution truth.

ChangeMesh separates authority into four lanes:

Deterministic Code
Gemini Semantic Judgment
Organizational Policy
Human Authority

This means a model cannot turn a failed deterministic test into a pass, and an execution agent cannot authorize its own actions.

Rehearse first with ShadowLab

One of the core ideas behind ChangeMesh is ShadowLab.

Before risky work progresses, ChangeMesh can rehearse the workflow against synthetic enterprise fixtures.

The rehearsal can test:

backward and forward compatibility;
failure recovery;
rollback behavior;
stale context;
policy boundaries;
agent restart and resume;
downstream breakage.

A simulation remains a simulation. ChangeMesh does not present rehearsal output as proof that a real external action occurred.

Human-on-the-loop with Approval Compression

Autonomous systems become much less useful if every step requires another approval.

ChangeMesh therefore uses Approval Compression.

Reversible and policy-permitted work can continue autonomously, while genuinely irreducible authority decisions are compressed into a bounded decision card containing the evidence, scope, rollback information, and exact authority being requested.

The goal is not to remove human authority.

It is to ask for it only where it is actually required.

Google Cloud runtime

ChangeMesh runs on Google Cloud.

The current architecture uses:

Google Cloud Run for the hosted application;
Firestore for durable workflow state;
Pub/Sub for asynchronous event flow;
Vertex AI and Gemini 3.6 Flash for bounded semantic reasoning;
Google ADK for multi-agent orchestration.

The system is designed around restart-safe, asynchronous workflows instead of assuming that an agent finishes everything in one chat session.

Evidence as a first-class output

A successful agent response is not enough.

ChangeMesh records deterministic evidence and produces a Change Evidence Passport so an operator or reviewer can see what actually happened.

The evidence model distinguishes different execution modes rather than collapsing them together.

For example, the current hosted ChangeMesh workflow is explicitly shown as SIMULATION.

Separately, the project contains historical LIVE_WRITE evidence from a bounded real GitHub Draft PR creation. That Draft PR was intentionally never merged.

The historical real action is not presented as the output of the current simulation.

That separation is intentional.

Testing and reproducibility

The frozen hackathon candidate includes more than 1,900 automated tests, reproducible setup instructions, an immutable release, a public Google Cloud deployment, an architecture diagram, and judge-facing evidence documentation.

The objective is simple:

Autonomous agents should not merely claim that a complex enterprise change is safe. They should be able to rehearse it, operate within bounded authority, and prove exactly what happened.