Spec-driven multi-agent architecture

Declare it once.
Run it anywhere.

A pyagent-blueprint YAML manifest compiles unmodified onto five structurally different runtime engines — validated, diffed, and contract-tested before a single LLM call.

Get started → GitHub pip install pyagent-all

customer-support.yaml

agents:
  classifier: {provider: fast}
  billing:    {provider: primary}
  tech:       {provider: primary}
workflows:
  support:
    pattern: supervisor
    agents:
      classifier: classifier
      routes: {billing: billing, tech: tech}

compile

pyagent Supervisor(classifier, routes={...}).run(input_)

langgraph StateGraph(...).add_node(...).add_edge(...)

crewai Agent(...) / Task(...) / Crew.kickoff_async()

openai_agents Agent(...) + Runner.run()

semantic_kernel Kernel() + ChatCompletionAgent.get_response()

Four independent pillars

Each one stands alone. None of them requires another.

Declare

Blueprint

Agents, workflows, providers, and governance in one reviewable YAML file — validate, diff, and test before running anything.

pyagent-blueprint Read more →

Execute

Execution & Routing

18 named orchestration patterns, difficulty-aware model routing, multi-provider fallback, and token budgets.

pyagent-patterns · router · providers · compress Read more →

Remember

Context & Memory

Three-tier memory with a trust-aware context ledger and PII redaction between agents.

pyagent-context Read more →

Observe

Observability

OpenTelemetry tracing, cost tracking, and a dashboard for replaying exactly what a run did.

pyagent-trace · pyagent-studio Read more →

Why not hand-write the orchestration instead

What a blueprint gets you that a script doesn't

Diff and review like infrastructure

pyagent-blueprint diff old.yaml new.yaml produces a semantic diff over the IR — not a text diff of hand-wired Python — so a reviewer sees exactly which agent, route, or SLA changed.

Governance is never silently dropped

Budgets, SLAs, memory tiers, and recovery policies are either honored by the runtime or surfaced as a stable diagnostic code like BUDGET_UNSUPPORTED — never silently ignored.

Validate before you run anything

Static analysis catches dangling references and schema violations before an LLM call is ever made.

Test without live API calls

pyagent-blueprint test checks contract conformance against a MockLLM, so CI validates a system's shape before spending a token.

Zero to running

Your first pattern, in Python — no YAML required

from pyagent_patterns.base import Agent, MockLLM
from pyagent_patterns.resolution import SelfReflection

pattern = SelfReflection(agent=Agent("coder", llm), max_rounds=3)
result = await pattern.run("Write an efficient Fibonacci function")

Where to start

Three doors in, depending on what you're doing

New to multi-agent systems

Start with Quickstart

Install, run your first pattern, then browse the 18-pattern catalog.

Have a real use case

Browse the Cookbook

37 complete, runnable recipes across 21 domains.

Evaluating against another framework

See the Adapters

The same manifest compiled onto LangGraph, CrewAI, and three other runtimes.