Open-Source Agent Orchestration

AI Agent Orchestration — With Guardrails

Version, replay, and debug every AI agent run — with circuit breakers that guarantee your agents never go rogue. Runs on your own hardware.

What happens when your AI agent goes rogue?

We gave an AI agent an unlimited research task and a $0.01 budget. Here’s what happened.

The $487 Problem

A developer posted their OpenAI bill: $487 in a single afternoon. Their autonomous research agent got stuck in a loop, making the same API call 2,000 times.

No budget cap. No kill switch. No circuit breaker. This is the default behavior of every AI agent framework today.

LLM API Costs (per 1M output tokens)

GPT-4o $10.00
Claude 3.5 Sonnet $15.00
DeepSeek V3 $1.10
Live — 6 events
$0.007 / $0.01
Research Overview #1 $0.001
Deep Dive: Technical #2 $0.001
Deep Dive: Risks #3 $0.001
Deep Dive: Market #4 $0.001
Cross-Analysis #5 $0.002
CIRCUIT BREAKER #6 $0.00
Circuit breaker triggered — Cost limit reached ($0.007/$0.01)

Flowmanner caught it in 3 seconds. Your budget is safe.

Three layers of protection

Every agent run is versioned, budgeted, and replayable.

Circuit Breakers

Set cost caps, time limits, and iteration budgets per agent. When a limit is hit, the agent stops — cleanly, immediately, no runaway costs.

Time-Travel Debugging

Every event logged. Click "Replay to here" on any timeline event to reconstruct the agent’s exact state. Git bisect for AI.

Auto-Assertions

After each run, 5 behavioral assertions are auto-generated: cost ceiling, latency, task completion, tool sequence, no circuit breaker.

Built for production AI agents

Run Diffing

Compare two runs side-by-side. See exactly where costs, tokens, and outcomes diverged.

Cost Pre-Flight

Before starting a run, see estimated cost with confidence interval. No surprises.

Sovereign Deployment

Runs on your own hardware. Your data never leaves your infrastructure.

Full Observability

Every LLM call, tool call, and decision logged as an append-only event stream.

Multi-Strategy Execution

Solo, DAG, pipeline, graph, swarm, and LangGraph strategies — pick the right one.

Budget Enforcement

Per-run cost caps with automatic shutdown. Your agents spend what you allow.

557
API Endpoints
7
Execution Strategies
67
Tests Passing
Apache 2.0
Open-Source Substrate

Tech Stack

React / Next.js Python / FastAPI PostgreSQL Redis Qdrant Docker WireGuard Jaeger
Early Access

Stop your next runaway agent

Join the waitlist. Be the first to know when Flowmanner opens to the public.

Read the Blog Post: How to Run AI Agents Without Going Bankrupt