THE PRODUCTION RUNTIME FOR AI AGENTS
Define a Pydantic schema and a prompt. Get a production HTTP backend with auth, typed outputs, pause/resume, and human approval. Deploy once. Call it a million times.
THE PATTERN BEHIND $20B+ IN AI VALUATIONS
Legora. Harvey. EvenUp. Basis. Billion-dollar valuations. Same core: an AI agent with domain knowledge, tools, and typed output. We open-sourced the engine.
Reads contracts
Legora
$5.5B
Legal assistant
Harvey
$11B
Demand letters
EvenUp
$2B
Automates accounting
Basis
$1.15B
WHAT THEY ALL HAVE IN COMMON
Pydantic enforces your schema. Invalid output? Auto-retry. No freeform text.
Agent asks a question. Waits days for the answer. Picks up with full context.
Side effects need sign-off. pending_actions → host approves → execute.
R2R + FastMCP. Upload PDFs. New regulation? New PDF, no code changes.
Pin to Anthropic for prompt cache. Fall back to others. 10x cost savings.
Open in Claude Code, Codex, Cursor, Gemini. Built-in skills teach the framework.
PRODUCTION-TESTED AT SCALE
Born from real enterprise work.
4M+
documents
$200k+
revenue
$0
VC raised
YOUR ENTIRE AGENT — 3 FILES, 30 LINES
# schemas.py
class ContractReview(BaseModel):
risk_level: Literal["low", "medium", "high"]
findings: list[Finding]
confidence: float
# agent.py
agent = create_agent(name="lawyer", output_type=ContractReview)
# main.py — that's it
app = create_app("lawyer")
# → Production API. Auth. Typed output. Done.
THIS IS NOT CHATGPT
CHATGPT
A chat window.
One user. Copy paste. No API. No guarantees. No scale.
AGENT2
Infrastructure.
Runs 24/7. Typed output. Human approval. Pause/resume. Call it like any API — except this one thinks.
Upload your knowledge. Define your schema. Deploy your agent. The production layer for AI agents shouldn't be a billion-dollar secret.
