One substrate, five layers.

Author — build with canvas or code, get both.

Agent workflows are authored on a visual canvas with full code parity: every graph is also a versioned, diffable definition. No lock-in, no drift between what the diagram shows and what runs.

nodes:
  trigger: inbound_email
  triage: agent.refund_bot
  gate:   gate.policy baseline-security
  pay:    tool.payments.issue_refund
edges:
  trigger -> triage -> gate -> pay
version: v1.3 (immutable)

WHAT'S UNDER IT

Graph DSL with keyboard-first canvas editor, git-friendly YAML export, immutable versions with a version pin per activation. Full schema in the DSL reference; versioning model in Graphs and versions.

Run — durable execution that survives anything.

Every step is checkpointed. A run survives infrastructure failure, and can pause at a human approval for weeks without holding compute. When the approver signs, the run resumes exactly where it stopped.

WHAT'S UNDER IT

Postgres-backed durable executor, Redis Streams lease queue, checkpoint every step, deterministic replay with non-determinism capture. Full detail — checkpoints, replay, time-travel fork — in Runs: durable execution.

Govern — policy on every edge, credentials on every agent.

Policy is evaluated on every edge traversal, tool call, credential use and model call — in-process, before the action reaches a system of record. Each agent is its own principal with scoped credentials rather than a shared key. Approvals carry quorum rules, service levels and escalation.

forbid (
  action in [Action::"invoke_tool"]
) when {
  resource.side_effects == "irreversible"
  && context.tainted
};

WHAT'S UNDER IT

Policy-as-code in Cedar, evaluated in-process by Cedar's Rust engine via cedarpy, scoped agent credentials via KMS, four-eyes approval gates with SLA + escalation, what-if simulation against past runs. See Policy packs and HITL gates and four-eyes.

Evidence — the record is the enforcement, not a report about it.

Every authorization, execution, approval and administrative action lands in a per-tenant hash-chained ledger with Ed25519-signed anchors. Evidence bundles export as a signed archive with a dependency-free verifier your auditor runs offline.

WHAT'S UNDER IT

Ed25519 signed anchors, per-tenant hash chain, offline verify.py bundle, atomic commit with the execution step. See The hash-chained ledger and Evidence bundles.

Improve — change only through tested updates.

No version reaches production without passing its regression evaluation suite. Overnight memory consolidation merges duplicate memories and surfaces contradictions — and anything it proposes stays inert until it passes the same suite and a human declines to veto it.

WHAT'S UNDER IT

Eval suites as first-class objects, promotion gated by regression eval, cross-run memory consolidation with human veto. See Evaluation suites and Memory consolidation.