Project 11 · Multi-Agent Systems · Async Architecture

AgentHandoff

Context Transfer Protocol

A durable state-transfer protocol that eliminates multi-agent "amnesia" — serializing full reasoning history to S3 and using SQS for guaranteed delivery between specialized AI agents in complex relay workflows.

View on GitHub Explore Architecture ↓

The Challenge

AI agents forget everything
when they pass the baton.

Multi-agent systems are only as powerful as their ability to coordinate. When one agent finishes and another begins, the reasoning history — the "why" behind every decision — is silently dropped by LLM context window limits.

The Problem

Multi-Agent Amnesia

When AI agents hand off tasks to one another in complex relay chains (Research → Code → QA → Deploy), the reasoning history and context accumulated by the first agent is truncated or lost entirely. The receiving agent starts blind — repeating work, making contradictory decisions, and breaking the chain of logic.

The Solution

Durable State Transfer

AgentHandoff serializes the complete agent context — reasoning chain, intermediate outputs, task metadata — to Amazon S3 as a structured JSON payload. Amazon SQS then guarantees delivery to the next agent in the chain with exactly-once semantics, full retry logic, and a complete audit trail of every handoff.

System Design

The Relay Race Protocol

A Research → Code → QA agent relay with zero context loss. Each agent receives the full reasoning history of every agent that came before it.

🔬

Research Agent

Gathers data, builds context payload

→

🗂️

Amazon S3

Serializes full context as JSON

→

📬

Amazon SQS

Guaranteed delivery + retry logic

→

⚡

Lambda Router

Resolves next agent, injects context

→

💻

Code Agent

Receives full research context

agenthandoff/handoff_protocol.py

# AgentHandoff — Durable context serialization
import boto3, json, uuid

s3 = boto3.client('s3')
sqs = boto3.client('sqs')

def handoff(agent_id, next_agent, context_payload):
  # Serialize full context to S3
  context_key = f"handoffs/{agent_id}/{uuid.uuid4()}.json"
  s3.put_object(
    Bucket='agent-context-store',
    Key=context_key,
    Body=json.dumps(context_payload)
  )

  # Queue handoff message for next agent
  sqs.send_message(
    QueueUrl=f"https://sqs.us-east-1.amazonaws.com/.../{next_agent}",
    MessageBody=json.dumps({
      'from_agent': agent_id,
      'context_s3_key': context_key,
      'task_id': context_payload['task_id']
    }),
    MessageGroupId=context_payload['task_id']
  )
  return {'status': 'handed_off', 'key': context_key}

Technology

Async Infrastructure
Built for Agent Relay Races

Every component is purpose-selected for durability and guaranteed delivery — the two non-negotiables for multi-agent coordination at scale.

Amazon S3

Durable context serialization store

Amazon SQS

Guaranteed message delivery & retry

AWS Lambda

Context injection & agent routing

Stateful Orchestration

Full reasoning chain preservation

Terraform

Infrastructure as Code

Python / Boto3

Lambda runtime & AWS SDK

Business Impact

Eliminates multi-agent amnesia.
Enables complex relay races.

AgentHandoff makes it possible to build AI workflows of arbitrary depth and complexity — with 100% context fidelity across every handoff.

🔗

Zero Context Loss

Every agent in the chain receives the complete reasoning history of all prior agents. No truncation, no summarization loss — full fidelity from first agent to last.

🏎️

Async Relay Chains

Enables complex Research → Code → QA → Deploy pipelines that run asynchronously at massive scale — without any agent waiting on another in real time.

📋

Full Audit Trails

Every handoff is logged with agent ID, timestamp, context key, and task ID. Complete observability into multi-agent workflows for debugging and compliance.