Embeddable workflow engine for AI agents. Deterministic, resumable, zero dependencies.

---

What is ForrestRun?

ForrestRun is a Python library for running AI agent workflows. Define your pipeline in YAML — mixing LLM agents, Python functions, and external tools — and run it with a few lines of code. Every step is persisted to SQLite, so you get replay, resume, and full state history out of the box.

from forrestrun import RuntimeBuilder

with RuntimeBuilder().with_model("openai/gpt-4o").with_db_path(":memory:").build() as runtime:
    run = runtime.run("workflows/research.yaml", inputs={"topic": "AI agents"})
    print(run.outputs)          # all step outputs
    print(run.total_tokens)     # token usage across all steps

Or for the simplest case — one function call, no setup:

from forrestrun import run_workflow

result = run_workflow("workflows/shopping.yaml", inputs={"shopping_list": "Buy a good book under 20 dollars"})
print(result.get_output("browse_and_checkout"))

Both import forrestrun and import agent_runtime work. No framework lock-in, no infrastructure, no dependency tree. pip install and go.

---

Why ForrestRun instead of LangGraph / CrewAI?

You want a library, not a platform. ForrestRun is a single pip install with two dependencies (PyYAML, typing-extensions). No HTTP client libraries, no vector databases, no required services. It embeds into your existing Python application — FastAPI, Django, scripts, Lambda — without pulling in an ecosystem.

You want determinism, not magic. Every step reads from and writes to an explicit state tree. No hidden state, no implicit globals, no action-at-a-distance. Workflows are YAML files you can diff, review, and version-control like any other config.

---

Define Workflows in YAML

Three step types — that's the whole model:

Step Type	What It Does	Backed By
type: agent	LLM reasoning — summarize, plan, decide, extract	Your agents/*.yaml
type: function	Deterministic Python — parse, classify, transform	Your functions/*.py
type: tool	External actions — HTTP calls, shell commands, file I/O	Your tools/*.py

schema_version: v1
workflow:
  id: research_and_act
  version: v1

inputs:
  topic:
    description: What to research

steps:
  - id: research
    type: agent
    agent: researcher
    inputs:
      topic: inputs.topic

  - id: classify
    type: function
    function: triage.classify_issue
    inputs:
      findings: steps.research.findings

  - id: notify
    type: tool
    tool: tools.http
    inputs:
      url: "https://hooks.example.com/alert"
      body: steps.classify.summary

---

What You Get

Feature	How It Works
SQLite-backed state	Every step's input, output, and state snapshot persisted atomically. Crash mid-run, lose nothing.
Deterministic replay	Re-run any past execution from stored state. No LLM calls, exact same output.
Resume from failure	Step 5 of 7 failed? Resume from step 5 — skip nothing, re-run nothing.
Conditional branching	Route execution based on runtime state: when: state.steps.classify.severity == "critical"
Multi-tier memory	Working, episodic, semantic memory layers available to agents across runs.
Full observability	Structured logs, HTML timeline visualization, step-by-step state inspection.
Zero HTTP dependencies	All LLM adapters use Python stdlib urllib. No httpx, no aiohttp.

---

Quick Start

pip install forrestrun

# scaffold a new project
mkdir my-agent && cd my-agent
ai init

# run a workflow
ai run workflows/example.yaml

LLM Providers

Configure in runtime.yaml — API keys resolved from environment or .env:

Provider	Models	Key
OpenAI	gpt-4o, gpt-4-turbo	OPENAI_API_KEY
Anthropic	claude-sonnet-4-20250514, claude-3-5-sonnet	ANTHROPIC_API_KEY
Gemini	gemini-2.5-flash, gemini-1.5-pro	GEMINI_API_KEY

All adapters use stdlib urllib and support structured multi-turn history for ReAct agents.

---

CLI

ForrestRun also ships a CLI for running, inspecting, and debugging workflows:

ai run workflows/example.yaml               # run a workflow
ai run workflows/example.yaml -i topic="AI"  # pass inputs
ai inspect <run_id> --steps                  # step-by-step breakdown
ai resume <run_id>                           # resume from failure point
ai replay <run_id> --verify-state            # deterministic replay
ai visualize <run_id>                        # HTML timeline
ai runs                                      # list recent runs

---

Project Structure

When you run ai init, ForrestRun scaffolds:

my-agent/
  agents/          # LLM agent definitions (YAML)
  functions/       # deterministic Python functions
  tools/           # custom tool implementations
  workflows/       # workflow definitions (YAML)
  runtime.yaml     # config (provider, model, limits)
  .env             # API keys

---

Docs

Topic	Link
Getting started	documentation/guide/getting-started.md
Writing workflows	documentation/guide/workflows.md
Writing agents	documentation/guide/writing-agents.md
Writing functions	documentation/guide/writing-functions.md
Writing tools	documentation/guide/writing-tools.md
Architecture	documentation/about/architecture.md
Full CLI reference	documentation/guide/cli-reference.md

---

Tests

pytest -q    # 635 tests

---

ForrestRun — the workflow engine that stays out of your way.