The open source backend for AI agents

Write a plain function; the backend turns it into a production endpoint with queues, retries, durable state, and versioned deploys. Build agents like APIs. Ship them like software.

Read the docs
apache-2.0python · typescript · goGitHub2.6k
forked and run in production at

If you can write an API, you can write an agent.

A reasoner is a plain function in Python, TypeScript, or Go. The plane turns it into a production endpoint your services call, with queues, retries, and streaming built in. The plane itself is a single lightweight Go binary that scales horizontally.

app = Agent("risk-engine")

@app.reasoner(tags=["fintech"])
async def score_transaction(tx: dict) -> dict:
    verdict = await app.ai(...)
    return verdict.model_dump()
POST /api/v1/execute/risk-engine.score_transactiondeployed

The parts you stop bolting on.

No Celery, no Redis, no Temporal, no separate vector store. The backend ships with the services you would otherwise run and glue together, each one a single call:

# the decorated function is already all of this:

$ curl -X POST .../execute/support.triage
{"status": "completed", "result": ...}

$ curl -X POST .../execute/async/support.triage
{"execution_id": "ex_9d2", "status": "queued"}
# poll it, or take the HMAC-signed webhook
every reasoner ships as sync and async endpoints with polling and signed webhooks · no gateway, no queue setup, no glue

From one agent to a fleet of thousands.

Nodes deploy independently; the plane routes, queues, retries, and scales. Executions stay durable for minutes or days. No timeout ceilings, no orchestration code.

1000s
concurrent agents on one plane
days
durable executions, not request timeouts
1
Go binary is the whole plane
3
SDKs: python, typescript, go

Deploy agents the way you deploy software.

Canary a new version on live traffic, run two reasoners head to head, switch blue to green in one step, shift load at runtime, let unhealthy versions drop out of routing on their own, and watch every execution stream through one pane.

v1 · 95%v2 · 5% canary
Canary releasesship the new agent to 5% of live traffic first
reasoner areasoner b
A/B at fleet leveltwo reasoners, same traffic, judged on results
blue · draininggreen · takes the load
Blue-greenthe load switches lanes instantly, zero downtime
Dynamic traffic controlshift load between agents at runtime, no redeploy
v3 · degraded, unrouted
Self-healing fleetunhealthy versions drop out of routing on their own
Observabilitymillions of executions, traced, streamed, searchable

When security asks, you already have the answer.

Every agent carries a cryptographic identity, every cross-agent call is signed, and every execution leaves a verifiable receipt. Your POC clears review without a meeting. The full governance story lives on the enterprise page.

$ af vc verify run-8f31.json
  signature   ed25519 valid
  chain       platform → node → function intact
  execution   risk-engine.score_transaction
offline verification, no agentfield service required verified

What teams build on it.

Any system that reasons and acts, on one plane. The gold entries ship as open blueprints: complete multi-agent systems you fork and adapt instead of starting blank.

fraud detectionfintech
risk scoringfintech
deep researchblueprint
customer triagesupport
voice agentsvoice
document intelligenceblueprint
clinical trial matchinghealthcare
supply chain risklogistics
security auditblueprint
sales intelligencerevenue
code reviewblueprint
yoursnext

Wired into the stack you already run.

Integrations are governed surfaces, not connectors. A warehouse event, an incident, or a pull request wakes an agent; its actions flow back out under the same identity, policy, and audit.

wired today
SnowflakeDatabricksGitHubSlackStripeAnthropicOpenAIGeminiOpenCode
early access
ServiceNow
on the roadmap
SalesforceHubSpotLinearJiraPagerDutyDatadogNotionGoogle SheetsTwilioOktaGitLabPostgresMongoDBRedisKafkaGoogle CloudCloudflareSentryLookerDiscordZoomAuth0AsanaAirtableConfluenceSAPMuleSoftCiscoSplunkShopifyBoxMistralHugging FaceZapier
every integration ships astriggeragent capabilityskill pack
See every AgentField integration

One install, laptop to production.

From the first function to a governed fleet, this is the only piece of infrastructure you add. Everything else you already have: your services, your data, and an inference provider. Open models on your GPUs or any API, swapped in one line. Apache-2.0, yours to fork and keep.

Qwen

model: qwen

The model is the interchangeable part. Every agent runs on the open model you point it at, and swapping is one line of config. No contract, no lock-in, no meter running against someone else's cloud.

Forked, extended, and run in production.

Developed in the open; forked and extended inside banks, software giants, universities, and public health institutions.

JPMorgan
ServiceNow
NIHnational institutes of health
SAP
Microsoft
CDW
USPsão paulo
MMinnesota
PSIOG
partners

Before you build

What engineers ask before they wire the first agent in.

ask on discord
Is it another agent framework?
No. It is infrastructure: a stateless Go control plane your agents register with. Write the agent in Python, TypeScript or Go.
How does anything call my agent?
Over plain REST. Every reasoner becomes POST /api/v1/execute/<agent>.<reasoner>, so a fetch() from the frontend works without an SDK.
What if a server dies mid-run?
The run is durable and resumes where it stopped. Long jobs run async for hours or days and report back by webhook.
Which models?
Yours. Open weights, a local endpoint or a hosted API, called on your key. The plane does not care which.
Do I need Redis or a vector database?
No. Memory is built in, scoped to a workflow, session or user, with vector search included.
How do I prove what an agent did?
Every agent carries a W3C DID, and every step emits a signed Verifiable Credential you can hand to an auditor.
How do I watch it run?
Each run draws a live execution DAG. Prometheus metrics sit at /metrics, and the logs are structured and correlated.
Where does it run?
On a laptop for the POC, in Docker or on your own cluster in production. Apache-2.0, inside your network.

Join the community

Ask how others wire their agents, share what you shipped, and hear about releases first.

Join the Discord

Taking it past the POC?

Identity, audit, private deployment, and a design partnership. The enterprise page makes the case to your leadership for you.

See enterprise

Ship your first agent as an API.

Install the backend, write a function, and call it over REST before lunch.

Quickstart