Features

• 🤝 Tool calls & handoffs — Agents consult each other or transfer control
• 🕸️ Pre-built topologies — Pipeline, Star, Mesh, or fully custom graphs
• 🧠 Shared agent memory — Persistent context across agents during a run
• 🗺️ Topology visualization — See your agent graph as an image
• 🐍 Pydantic-AI Native — Use Pydantic-AI and Logfire observability
• ⚙️ Configurable context passing — Control what flows between agents

Installation

pip install pydantic-collab

🚀 Quick Start

from pydantic_collab import PipelineCollab, CollabAgent

collab = PipelineCollab(
    agents=[
        CollabAgent(name="Triager", system_prompt="Classify the support ticket by product area and urgency (P0-P3). Extract the core issue."),
        CollabAgent(name="Responder", system_prompt="Draft a helpful response. Acknowledge the issue, provide next steps or workarounds."),
    ],
    model="anthropic:claude-sonnet-4-5",
)

result = collab.run_sync("User email: 'I can't export my data to CSV, the button just spins forever. I need this for a board meeting tomorrow!'")
print(result.output)

When to Use This

Use pydantic-collab when you need multiple specialized agents working together, without the need of human intervention.

Relevant Use Cases:

• Multi-stage workflows (triage → investigate → respond)
• Specialist teams (coordinator + domain experts)
• Tasks requiring different perspectives (engineering + legal + security)
• Iterative refinement loops (drafter ↔ critic)

⚒️ 🤝 Tool Calls vs Handoffs

Common Topologies

Pipeline (Sequential)

from pydantic_collab import PipelineCollab, CollabAgent

# Incident postmortem pipeline
collab = PipelineCollab(
    agents=[
        CollabAgent(name="TimelineBuilder", system_prompt="Construct a timeline from the incident channel logs. List each event with timestamp and actor."),
        CollabAgent(name="RootCauseAnalyzer", system_prompt="Identify contributing factors and the root cause. Distinguish symptoms from causes."),
        CollabAgent(name="ActionItemWriter", system_prompt="Propose concrete, assignable action items with owners. Prioritize by impact."),
    ],
    model="openai:gpt-5.2-pro"
)

# Use: collab.run_sync(slack_channel_export)

Star (Hub & Spoke)

from pydantic_collab import StarCollab, CollabAgent

# Feature spec review - PM coordinates specialists
collab = StarCollab(
    agents=[
        CollabAgent(name="ProductLead", system_prompt="Coordinate the review. Synthesize feedback from specialists into a go/no-go recommendation."),
        CollabAgent(name="EngineeringReviewer", system_prompt="Assess technical feasibility, estimate complexity, flag architectural concerns."),
        CollabAgent(name="DesignReviewer", system_prompt="Evaluate UX implications, accessibility, consistency with design system."),
        CollabAgent(name="SupportReviewer", system_prompt="Predict support burden, documentation needs, and customer confusion risks."),
    ],
    model="openai:gpt-5.2-pro"
)

# ProductLead consults each specialist, then synthesizes

Mesh (Everyone talks to everyone)

from pydantic_collab import MeshCollab, CollabAgent

# Vendor evaluation - each perspective needs input from others
collab = MeshCollab(
    agents=[
        CollabAgent(name="SecurityReviewer", system_prompt="Assess security posture: SOC2, data handling, access controls. Consult Legal on compliance implications."),
        CollabAgent(name="EngineeringEvaluator", system_prompt="Evaluate integration effort, API quality, scalability. Consult Security on auth requirements."),
        CollabAgent(name="LegalReviewer", system_prompt="Review contract terms, data processing agreements, liability. Consult Security on data residency."),
    ],
    model="openai:gpt-5.2-pro"
)

# Each agent can consult others - security implications affect legal review, etc.
# Use: collab.run_sync("Evaluate Acme Corp's proposal for our analytics pipeline: ...")

from pydantic_collab import Collab, CollabAgent

# Code review with specialist consultation
collab = Collab(
    agents=[
        CollabAgent(
            name="LeadReviewer",
            system_prompt="Review the PR for correctness and design. Consult specialists for deep dives.",
            agent_calls=["SecurityChecker", "PerfAnalyzer"],  # Can consult
            agent_handoffs="SummaryWriter",                   # Hands off for final summary
        ),
        CollabAgent(name="SecurityChecker", system_prompt="Check for vulnerabilities: injection, auth issues, secrets."),
        CollabAgent(name="PerfAnalyzer", system_prompt="Identify performance issues: N+1 queries, blocking calls, memory."),
        CollabAgent(name="SummaryWriter", system_prompt="Compile all feedback into a clear, actionable review summary."),
    ],
    model="openai:gpt-5.2-pro",
   final_agent="SummaryWriter"
)

🧠 Agent Memory

Share persistent context between agents during a run:

from pydantic_collab import PipelineCollab, CollabAgent, AgentMemory

incident_context = AgentMemory(
    name="incident",
    description="Accumulated incident context: affected systems, customer impact, timeline"
)

collab = PipelineCollab(
    agents=[
        CollabAgent(
            name="FirstResponder",
            system_prompt="Gather initial context: what's broken, who's affected, when it started. Document in memory.",
            memory={incident_context: "rw"},  # Writes initial context
        ),
        CollabAgent(
            name="Investigator",
            system_prompt="Deep dive into the root cause using the established context. Add findings to memory.",
            memory={incident_context: "rw"},  # Reads context, adds findings
        ),
        CollabAgent(
            name="CommunicationDrafter",
            system_prompt="Draft customer and internal communications based on the full incident context.",
            memory={incident_context: "r"},   # Reads accumulated context
        ),
    ],
    model="openai:gpt-5.2-pro"
)

# Simple string (defaults to 'rw')
CollabAgent(name="Agent", memory="notes")

# List (all default to 'rw')
CollabAgent(name="Agent", memory=["notes", "decisions"])

# Dict for explicit permissions
CollabAgent(name="Agent", memory={"notes": "rw", "config": "r"})

Visualizing Topology

collab = Collab(...)
collab.visualize_topology()  # Opens image
collab.visualize_topology(save_path="topology.png", show=False)  # Save to file

pip install pydantic-collab[viz]  # Requires visualization dependencies

Adding Tools

collab = Collab(agents=[...], model="openai:gpt-4o-mini")

@collab.tool_plain
async def query_metrics(service: str, time_range: str) -> str:
    """Query Prometheus metrics for a service."""
    # ... integration with your metrics system
    return metrics_data

@collab.tool_plain(agents=("Investigator",))  # Only for specific agents
async def get_recent_deploys(service: str) -> str:
    """Get recent deployments for a service."""
    return deployment_history

Result Object

result = collab.run_sync("Query")

result.output              # Final output
result.final_agent         # Agent that produced output
result.execution_path      # ["Triager", "Investigator", "Responder"]
result.usage               # Token usage statistics

print(result.print_execution_flow())  # Visual flow diagram

Configuration

collab = Collab(
    agents=[...],
    max_handoffs=10,           # Maximum handoff iterations (default: 10)
    max_agent_call_depth=3,    # Maximum recursive tool call depth (default: 3)
)

from pydantic_collab import CollabSettings

collab = Collab(
    agents=[...],
    collab_settings=CollabSettings(
        include_conversation="allow",      # "allow", "disallow", "force"
        include_thinking="disallow",
        include_handoff="allow",
        include_topology_in_prompt=True,
    ),
)

from pydantic_collab import PromptBuilderContext, CollabSettings

def my_prompt_builder(ctx: PromptBuilderContext) -> str:
    lines = [f"Agent: {ctx.agent.name}"]
    if ctx.can_handoff:
        lines.append(f"Hand off to: {', '.join(a.name for a in ctx.handoff_agents)}")
    return "\n".join(lines)

collab = Collab(
    agents=[...],
    collab_settings=CollabSettings(prompt_builder=my_prompt_builder),
)

from pydantic import BaseModel

class MyDeps(BaseModel):
    db: Database
    cache: Cache

collab = Collab(agents=[...])
result = collab.run_sync("...", deps=MyDeps(db=db, cache=cache))

Examples

See examples/ for complete working examples:

uv run --env-file .env examples/01_simple_chain.py

License

This project is licensed under the MIT License - see the LICENSE file for details.