Flowgent

Predictable, auditable, distributed enterprise AI agent orchestration — with native economic layer (x402/MPP).

Flowgent is an AI-native universal orchestration engine modeled after Apache Flink's session/application architecture. It deeply integrates LLM agent intelligence with deterministic DAG execution — preserving the predictability, reliability, and auditability of traditional workflows while empowering distributed enterprise super-agents with bounded, governable autonomy.

---

Features

• DAG Topological Scheduler — predictable execution with concurrent fan-out. 11 node types compose any orchestration topology.
• Deterministic + Intelligent — agent and supervisor nodes provide LLM intelligence; all others guarantee deterministic execution.
• Controlled Autonomy — supervisor constrained to continue|retry|inject|abort with configurable quotas.
• State-Machine Persistence — every run and task is durable. Pause at human gates, resume via API, replay idempotently.
• Multi-Namespace API — namespace-scoped REST paths (/api/v1/{namespace}/...), JWT/OIDC/GitHub OAuth, A2A protocol server (Google Agent-to-Agent).
• Application Mode — every agentflow gets a dedicated per-flow K8s JobManager, auto-scaling TaskManagers/sandboxes. (Session mode — a shared JM/TM pool — is temporarily disabled to simplify troubleshooting; the priority field is reserved for its reintroduction, but high is the only value currently accepted.)
• Dual-Mode Deployment — All-in-One (SQLite + memory queue) or Production (PostgreSQL + MQTT/EMQX + Redis + K8s).
• OTEL Tracing Per Node — every node span records input, output, and internal state for Jaeger debugging.
• Cron + Webhook Triggers — schedule-based and event-driven (GitHub/GitLab webhook) per agentflow.
• Multi-Provider LLM — OpenAI-compatible adapter with per-provider rate limiting, SOCKS/HTTP proxy, modalities, and extended thinking.
• JSON Schema Validation — agent output validated against output_schema with auto-retry on failure.
• Notification Service — Telegram, DingTalk, Slack, Email, Webhook channels; WebSocket SSE push for human approvals.
• Optional Economic Layer — x402 payment protocol with spending policies, wallet abstraction, and Coinbase facilitator integration.

---

Quick Start

Prerequisites: Go 1.26+ · (optional) PostgreSQL 15+, EMQX 5.x, Redis 7.x

git clone git@github.com:flowgent-labs/flowgent.git && cd flowgent
make build:image     # build Docker images (core + wallet)
make build            # build all binaries (core + wallet + MCPs)

• Run with a config file (required):

# Development (SQLite + memory queue)
./bin/flowgent daemon start -c etc/flowgent.yaml

# Use the fully annotated sample as a starting point
cp etc/flowgent.yaml my-config.yaml
./bin/flowgent daemon start -c my-config.yaml

• Verify

curl http://localhost:9999/_/healthz             # {"status":"ok"}
curl http://localhost:9999/_/openapi.yaml        # OpenAPI 3.1 spec
curl http://localhost:9992/.well-known/agent.json # A2A agent card

REST API on :9999 · A2A on :9992 · Wallet on :9901 · pprof on :9991

---

Architectures

• docs/01-L1-Engine-Architecture.md
• docs/02-L1-x402-Economic-Support.md

Examples

• docs/10-L2-USE-CASES.md

---

License

See Apache License.