Build in your framework. Govern with us. Run on your model.
Three moves, and none of them asks you to abandon the tools your engineers already chose.
None of them asks you to start over.
Build in your framework
LangGraph, the OpenAI Agents SDK, whatever your engineers already chose. We never compete on authoring ergonomics — we govern whatever they produce.
agent = LangGraph(...) # unchanged Govern with AgentOS
A thin SDK over open contracts gives the agent an identity, scoped credentials, delegated authority and an audit trail. It stays a lean standalone service.
from agentos import principal, policy Run on your model
Model is a configuration value. Cloud is any Kubernetes. Identity federates to your IdP. Leave whenever you like — that is the point.
LLM_PROVIDER=bedrock # no rebuild Every action walks the same path
Reasoning is delegated to the agent. Execution flows through a deterministic, recordable path. This single principle is why we can promise that even at thousands of concurrent agents you can reconstruct who initiated an action, why a transition occurred, and which policy allowed or blocked it.
| Hop | What happens | Result |
|---|---|---|
| 01 Identity & delegation | Agent presents a subject/actor chain, verified at the edge | passed |
| 02 Policy decision | Re-derived from the store at request time — never from token claims | passed |
| 03 Human approval | Durable primitive; the workflow waits days if it must | passed |
| 04 Durable orchestration | The workflow replays exactly where it left off | evaluating |
| 05 Audit event | Correlated, queryable, reconstructable across thousands of agents | queued |
finance-recon → post journal entry · $84,200
A real action shape, cycling through the five governed hops.
✓ approved by m.okonkwo · 00:04:12 elapsed
✓ audit emitted · 14 hops · 0 unscoped calls
✗ egress gate: provenance ∉ requester entitlement — denied An approval can wait days without losing its place
Human-in-the-loop approval is a platform primitive, not application code. A task interrupted mid-flight — even mid-approval, days in — resumes exactly where it left off.
Cost is not just observed, it is a control
Per-action token and cost capture is attributed to worker, person, team and model supply channel — then used to enforce windowed quotas, budget envelopes on delegated work, and deterministic routing that falls back across supply channels instead of silently downgrading models.
Restarts do not lose the thread
The workflow replays exactly where it left off. Minutes or days, across process restarts, the approval keeps its place and the audit trail stays correlated.