Introduction

Flightdeck is an open-source AI agent orchestrator that makes it easy to build and compose without worrying about states and infrastructure.

Why Flightdeck?

Flightdeck is an framework layer purpose-built for running AI agents in production. It addresses four critical areas:

Production agent systems need to handle failure gracefully and maintain consistent state. Flightdeck provides:

State management — Persistent, checkpointed execution state so agents can resume from where they left off after interruptions
Scalability — Horizontal scaling of agent workloads with built-in queue management and resource-aware scheduling
Fault tolerance — Automatic retries with exponential backoff, circuit breakers for downstream services, and graceful degradation when individual agents fail

Running agents without visibility is flying blind. Flightdeck gives you full operational insight:

Cost control — Per-agent and per-task token usage tracking with budget limits and alerts to prevent runaway spend
Auditing — Complete execution traces capturing every LLM call, tool invocation, and decision point for compliance and debugging
Latency monitoring — End-to-end timing for tasks, tool calls, and LLM round-trips with bottleneck identification and SLA tracking
Logging — Structured, leveled logs across all agent activity with correlation IDs to trace requests end-to-end through multi-agent workflows
User feedback — Built-in feedback collection hooks that tie end-user ratings and corrections back to specific agent runs for continuous improvement

Flightdeck is designed to get out of your way so you can focus on what your agents actually do:

No complex abstractions — Business logic and infrastructure are cleanly separated. Define what your agents do in plain, straightforward code while Flightdeck handles execution, scaling, and recovery behind the scenes
Language agnostic — First-class SDKs for Python, TypeScript, and Go. Define agents in whichever language your team already uses, and mix languages within the same workflow
Unit testing — Test agents, tools, and workflows in isolation with built-in mocks for LLM calls and tool responses. Validate behavior deterministically without burning tokens or hitting external services