Architecture

General Augment hosts and governs customer Agents; Hermes runs them.

CLI or dashboard
  -> load one Project desired-state head
  -> apply one typed mutation and semantic plan
  -> atomic revision + Candidate + plan + operation + audit + outbox
  -> leased reconciliation and provider verification
  -> hosted Agent evidence
  -> exact owner/admin review
  -> Live pointer

The dashboard and CLI call the same v3 APIs. A dashboard save does not update mutable Agent rows or invoke a separate release compiler.

Runtime

authenticated app backend
  -> POST /v1/responses with Project target key, Agent, and user
  -> memory/Redis target-pointer cache
  -> immutable execution-plan cache by release fingerprint and Agent ID
  -> Hermes turn
  -> exact tools, memory, limits, and policy
  -> response, trace, usage, and audit

A cache hit performs no Project graph reconstruction or configuration-table fan-out. Effective policy is the Project ceiling intersected with the Agent narrowing.

Host app

The app keeps its UI, auth, data, billing, and business logic. Its server holds the General Augment runtime key, passes a stable authenticated user identity, and verifies signed direct-tool callbacks. Browser and mobile clients never hold runtime or provider credentials.

New apps, add-Agent projects, and agent-backend replacements use the same issuer/JWKS, user claim, callback, environment, and connection contract. Coding-agent adapters modify only the host app.

Storage and workers

PostgreSQL stores Workspaces, Project source metadata, immutable revisions/releases, operational Project data, observed state, audit, and the transactional outbox.
GCS stores content-addressed source bytes in production; local development uses filesystem assets.
Redis holds target and execution-plan caches, queues, leases, rate limits, and hot counters.
ARQ workers reconcile actual external state and generate hosted evidence after the durable receipt returns.
The auth proxy keeps credentials outside Hermes context and injects them only at execution.

See the generated API reference for exact routes and schemas.