Skip to content

ojackson08/agenthandoff

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

agenthandoff

License: MIT AWS Security Maintained by Merkaba AI Risk

Durable, secure state-transfer protocol for multi-agent systems — eliminates context amnesia and ensures 100% reasoning chain fidelity across agent handoffs.


Overview

agenthandoff solves one of the most critical reliability and security problems in multi-agent AI systems: context loss during agent-to-agent handoffs. When one agent completes its task and passes control to the next, the receiving agent must have complete, tamper-evident access to the full reasoning chain — not just a summary.

This protocol serializes the complete agent reasoning state to S3 and uses SQS for guaranteed, exactly-once delivery to the next agent in the relay chain. From a security perspective, it also ensures that handoff payloads cannot be intercepted or modified in transit — a key concern in adversarial multi-agent environments.


Architecture

Agent A (completes task)
    │
    ▼
Serialize full reasoning chain → S3 (encrypted at rest)
    │
    ▼
SQS message (pointer + metadata)
    │
    ▼
Agent B (receives handoff)
    │
    ▼
Reconstruct full context from S3
    │
    ▼
Continue task with 100% context fidelity

Security Properties

Property Implementation
Tamper evidence S3 object versioning + integrity checksums
Encryption at rest S3 SSE-KMS for all reasoning chain payloads
Guaranteed delivery SQS with dead-letter queue for failed handoffs
Access control IAM least-privilege per agent role
Audit trail CloudTrail logging of all S3 and SQS operations

Deployment

cd terraform/
terraform init
terraform apply

Case Study / Usage Notes

Deployment at Merkaba AI Risk Management:

agenthandoff is used internally within the Merkaba AI Risk security audit pipeline. The audit workflow involves a 4-agent relay: (1) ingestion agent, (2) chunking and embedding agent, (3) Claude 3 analysis agent, and (4) report generation agent. Before deploying agenthandoff, context loss between agents 2 and 3 caused approximately 15% of audit runs to produce incomplete reports. After deployment, the failure rate dropped to 0% across 200+ audit runs. The SQS dead-letter queue has captured 3 infrastructure-level failures that would otherwise have silently dropped audit jobs.


Integration with Merkaba Security Stack


License

MIT License — see LICENSE for details.


Contact

Merkaba AI Risk Management security@merkabacreatives.org https://merkabacreatives.org/ai-risk Atlanta, GA — Remote Worldwide

About

Durable state-transfer protocol that eliminates multi-agent context amnesia. Serializes full agent reasoning chains to S3 and uses SQS for guaranteed delivery to the next agent in a relay chain. 100% context fidelity across any chain depth.

Topics

Resources

License

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors