Skip to content

The requester can change.The execution path does not.

Coding agents are good at expressing intent against typed interfaces. Give them an environment to ask, not a pile of YAML to reverse engineer.

agent sessionThin UX

A coding agent asks what depends on orders-db, then proposes an upgrade that goes through plan, policy, and approval

agent> what depends on orders-db in prod?
orders-db  aws-rds-postgres@1.3
├── billing-api
├── worker
└── reporting-job

agent> propose upgrading orders-db to 1.4
plan     2 changes, 0 destroys
impact   3 dependents, connection params unchanged
policy   ✓ passed (prod-db-change)
approval required from @platform
→ proposal #418 waiting for review

Agents get the same environment context humans do.

An agent can understand applications, infrastructure, dependencies, versions, and change impact without reverse-engineering deployment pipelines or receiving unrestricted cloud credentials.

  • pipeline.yml × 14 repos
    How does billing-api reach production?
    environment graph
  • AWS admin access keys
    What am I allowed to do here?
    propose only · the execution worker holds the credential
  • terraform/ across 37 repos
    What infrastructure does this application depend on?
    orders-db: aws-rds-postgres@1.3
  • approval rules in a wiki
    Does this change require approval?
    prod-db-change requires @platform

Ask the environment

The dependency graph is data. An agent can ask what depends on orders-db, which versions are deployed where, or what a change would touch, and get a typed answer.

It is the same picture a developer sees and a platform engineer reviews against, so the agent works from the architecture everyone else is looking at.

agent>what depends on orders-db?orders-dbaws-rds-postgres · 1.3 · prodbilling-apilambda · 1.8workerecs-service · 1.2reporting-jobk8s-cronjob · 0.73 dependents · billing-api, worker, reporting-job

Propose, do not apply

An agent’s change becomes a proposed deployment with a plan and an impact report. Policy runs on it. If the environment requires approval, a person approves it. The agent never holds the cloud credential.

THE AGENT STOPS AT THE PROPOSALagentpropose orders-db 1.3 → 1.4 --env prodplan + impact3 dependents · 0 destroyspolicyprod-db-change ✓approvalwaiting for @platformapproved by @platformexecution workerTerraform · OpenTofu · Helmprodorders-db 1.4 · change #418 recordedcloud credential

Interfaces agents can use today

MCP serverToday
Claude Code and other MCP clients

Query environments, list dependencies, propose and approve deployments.

CLIToday
any agent with a shell

Every command runs with the permissions of the identity it runs as.

GraphQL APIToday
your own automation

The API the UI and CLI use, scoped by service accounts and ABAC policy.

One path into the environmentpermissions · policy · approvals on every request

Give your agent an environment