How Beck Works
Beck reasons against the real schema, the real code and the live state of your infrastructure. Three layers make this possible: a squad of agents, a tool layer and the state ledger.
A request reaches a coordinator, which routes it to a squad of agents. The squad uses MCP tools and reads the live state from the Spanner ledger. The result is a pull request that a person can read and merge.
Agent squad
Beck is a multi-agent system. A coordinator sends each request to the specialist that fits it:
- a business analyst, who turns a request into requirements;
- software and enterprise architects, who shape the solution;
- platform and devops engineers, who write the manifests;
- a QA evaluator, who checks the output before Beck proposes it.
Sub-agents do the focused work: a portal assistant, a manifests author, a GitHub integrator, an analyst for logs and metrics, and a knowledge base.
Tools
Beck reaches outside systems through Model Context Protocol (MCP) tools.
| Source | What it gives |
|---|---|
infrastream-mcp | The live state and dependency graph from the Spanner ledger. Beck queries it. It holds no copy. |
code-graph-mcp | A semantic graph of your source code. |
| Knowledge base | The manifest schema, the Manifest Reference and the samples. |
| Google Cloud signals | Live metrics, traces and logs. |
State ledger
Infrastream records each resource in Spanner with fields that exist for the agent:
- Semantic embeddings of the configuration and description, so Beck can find "databases like this one" by meaning.
- Search context and links for each resource.
- Root-cause and error attribution on the state itself, so a failure carries a structured reason.
- Topology markers that show how resources cluster.
The state is a directed dependency graph. It is the graph that the engine executes. Beck follows it in both directions to find what a change would break. It respects order: the APIs and IAM of a project must exist before the services that use them.
Beck holds the desired state (your manifests), the actual state (what the engine built) and the run history. For a failure, it joins the run events and the graph with live logs, metrics and traces. It then names the IAM binding, the quota or the dependency that does not exist.
Data boundaries
Beck reasons over infrastructure state and metadata. It works inside the permission and tenancy limits of your organization. It does not read the private data of your applications. The changes it proposes arrive as pull requests. Your branch rules set how strictly the team reviews them.