Architect 2.0: proposed production architecture

Architect 2.0: proposed production architecture An architecture diagram generated by Archify. GitHub App · fine grained, per repo · Architecture component GitHub App fine grained, per repo Sandbox pool · one microVM per build · Build plane: one sandbox per build, egress allowlisted Sandbox pool one microVM per build Secrets vault · never in model context · Build plane: one sandbox per build, egress allowlisted Secrets vault never in model context AI gateway · routing, BYOK, metering · Egress control for builds AI gateway routing, BYOK, metering Builder workspace · chat, plan, agents, code · Client, untrusted Builder workspace chat, plan, agents, code API tier · stateless, autoscaled · Platform control plane, stateless and autoscaled API tier stateless, autoscaled Build job queue · durable, resumable · Platform control plane, stateless and autoscaled Build job queue durable, resumable Agent harness · plan, then edit, run, test · Platform control plane, stateless and autoscaled Agent harness plan, then edit, run, test Model providers · Jev, language, code · Architecture component Model providers Jev, language, code Postgres · projects, versions, meters · Tenant data · PostgreSQL Postgres projects, versions, meters Object store · snapshots, logs, traces · Tenant data Object store snapshots, logs, traces Preview and production · a preview per version · Run plane: deployed apps and agents, its own proxy Preview and production a preview per version Runtime AI proxy · for deployed apps · Run plane: deployed apps and agents, its own proxy Runtime AI proxy for deployed apps HTTPS after consent webhooks tenant scoped one durable job one worker typed tools at run time every step by task type snapshot per step promote on health agent calls no provider key Client, untrusted Platform control plane, stateless and autoscaled Build plane: one sandbox per build, egress allowlisted Egress control for builds Tenant data Run plane: deployed apps and agents, its own proxy

What the shape buys a builder

  • • Per stage meters at the gateway are where a cost estimate comes from, before the build rather than after it
  • • A trace per step is what answers "Why did it do that?" without reading a log
  • • A checkpoint commit per step, so a bad step is one revert rather than a rebuild

Boundaries that do real work

  • • Secrets reach the sandbox at run time and never enter model context
  • • One sandbox per build with an egress allowlist, so a generated command cannot reach the platform
  • • Deployed apps hold no provider keys: they call agents through the runtime proxy

Proposed, not built

  • • This is the target architecture, written up in docs/ARCHITECTURE.md
  • • The demo implements sign in, per user Postgres rows and live decision model calls
  • • What runs today is the separate system architecture diagram