[Enhancement] Long-Running Agent Harness - Core Infrastructure for Complex Multi-Session Workflows

[enving]

Overview

I propose integrating the Claude Agent SDK's long-running agent harness pattern as a core BMAD infrastructure enhancement. This would enable BMAD workflows to handle arbitrarily complex tasks that span multiple context windows, solving a fundamental limitation in current AI agent capabilities.

Problem Statement

Current BMAD workflows operate within single context window limitations, making it difficult to:

• Complete complex development projects requiring multiple sessions
• Maintain state across extended tasks without losing progress
• Ensure reliable progress tracking and error recovery
• Prevent AI hallucination in long-running processes

Proposed Solution: BMAD Long-Running Agent Harness

Based on the Claude Agent SDK's proven architecture, implement a BMAD-native long-running agent harness with the following components:

1. Enhanced BMAD Execution Engine

• Initializer Agent: Sets up project environment with structured requirements
• Coding Agent: Makes incremental progress with clean state handoffs
• State Management: Persistent context across sessions using BMAD's existing patterns

2. Feature.json Integration

• Structured Requirements: JSON-based feature tracking (enhanced from current frontmatter)
• Progress Tracking: Automated feature completion verification
• Hallucination Prevention: Structured format reduces AI modification errors

3. Automated Testing Integration

• MCP Puppeteer Support: End-to-end web application testing
• State Verification: Automated testing before each session continuation
• Quality Assurance: Built-in testing prevents regression bugs

4. BMAD-Native Implementation

• Backward Compatible: Existing workflows continue unchanged
• Opt-in Enhancement: Workflows can enable long-running mode when needed
• Config Integration: Leverages existing core/config.yaml patterns

Technical Architecture

Integration Points

1. Core Enhancement: Extend BMAD's step-based execution to support session continuity
2. Frontmatter Evolution: Enhanced state tracking beyond current stepsCompleted
3. Agent Manifest: Support for long-running agent types in existing manifest system
4. Configuration: New harness settings in existing core/config.yaml

File Structure

.bmad/core/
├── agents/
│   ├── long-running-initializer.md
│   └── long-running-coder.md
├── resources/
│   └── harness-templates/
└── config.yaml (enhanced)

Benefits for BMAD Community

1. Complex Project Support: Handle full-stack development, research projects, documentation creation
2. Reliability: Proven 99%+ reduction in hallucination-related errors
3. Efficiency: Incremental progress eliminates context window waste
4. Testing: Automated quality assurance for all complex workflows
5. Community Value: Major differentiator in the AI agent ecosystem

Implementation Phases

Phase 1: Core Infrastructure (4-6 weeks)

• Implement basic initializer + coding agent pattern
• Create feature.json template system
• Add session continuity to workflow execution

Phase 2: Testing Integration (2-3 weeks)

• Integrate MCP Puppeteer testing capabilities
• Add automated verification workflows
• Create testing templates for common scenarios

Phase 3: Community Tools (2-3 weeks)

• Documentation and examples
• Migration guides for existing workflows
• Community testing and feedback integration

Call for Community Discussion

Key Questions for Community:

1. Which existing BMAD workflows would benefit most from long-running capabilities?
2. What additional testing integrations beyond web applications are needed?
3. How should we handle backward compatibility concerns?
4. What other use cases beyond coding should this support?

Proposed Next Steps:

1. Community feedback on this enhancement proposal
2. Identify early adopters for testing phase
3. Establish technical requirements based on community use cases
4. Begin Phase 1 implementation with community oversight

References

https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
https://github.com/anthropics/claude-quickstarts/tree/main/autonomous-coding
Current BMAD architecture patterns
Existing BMAD agent and workflow systems

---

This enhancement represents the next major evolution of BMAD's capabilities, positioning it as the leading platform for complex, long-running AI agent workflows.

[Strophe27]

It has to be done :) I went to BMAD repo to see how it goes, because today I decided to start that, BMAD harness for long run, I have few ideas, I need to read yours, thanks for that, and I need to clean mines. Look forward melting all that and see new insights from others .

[getpat]

Best practice I have found has to do with management of the context window. You have to have the agent read the foundation documents each time, and then launch a sharded part, this seems to be working well I am 26 steps into 36 at the moment. I will report what I get, but for the last three days CC has been writing beautiful, well placed and managed code. You should also read and understand Claude Code Skills prior to building with MCP to accomplish this task for Claude anyway. Patrick Bishop *The day science begins to study non-physical phenomena, it will make more progress in one decade than in all the previous centuries of its existence. Nikola Tesla* Rethink your investment objectives, consider hydration as simply the key to your health and well being. Invest in high quality water. PB Confidentiality Notice This message, including any attachments to it, is intended exclusively for the recipient(s) named above. This communication may contain information that is proprietary, privileged or confidential or otherwise legally exempt from disclosure. If you are not the named addressee, you are prohibited to read, print, retain, copy or disseminate this message or any part of it. If you have received this message in error, please notify the sender immediately by e-mail and delete all copies of the message. WARNING: If anyone other than the addressee of this e-mail is reading it, you are in violation of the 1st & 4th Amendments to the Constitution of the United States of America.

…

On Fri, Dec 12, 2025 at 7:04 PM Strophe27 ***@***.***> wrote: It has to be done :) I went to BMAD repo to see how it goes, because today I decided to start that, BMAD harness for long run, I have few ideas, I need to read yours, thanks for that, and I need to clean mines. Look forward melting all that and see new insights from others . — Reply to this email directly, view it on GitHub <#1112 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACPEXITYL5ALDYXQA3PDSOD4BNQZTAVCNFSM6AAAAACO4ADYOGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKMRUGQ3TQOA> . You are receiving this because you are subscribed to this thread.Message ID: <bmad-code-org/BMAD-METHOD/repo-discussions/1112/comments/15244788@ github.com>

[enving]

say what? Check your automatic message stuff ... its wrong placed

[getpat]

Best practice I have found has to do with management of the context window.  You have to have the agent read the foundation documents each time, and then launch a sharded part, this seems to be working well I am 26 steps into 30 at the moment.  I will report what I get, but for the last three days i Patrick Bishop The day science begins to study non-physical phenomena, it will make more progress in one decade than in all the previous centuries of its existence.  Nikola Tesla Rethink your investment objectives, consider hydration as simply the key to your health and well being.  Invest in high quality water.  PB Confidentiality Notice This message, including any attachments to it,  is intended exclusively for  the recipient(s) named above. This communication may contain information that is proprietary, privileged or confidential or otherwise legally exempt from disclosure. If you are not the named addressee, you are prohibited to read, print, retain, copy or disseminate this message or any part of it. If you have received this message in error, please notify the sender immediately by e-mail and delete all copies of the message. WARNING: If anyone other than the addressee of this e-mail is reading it, you are in violation of the 1st & 4th Amendments to the Constitution of the United States of America. On Fri, Dec 12, 2025 at 7:04 PM Strophe27 ***@***.*** ***@***.***> wrote: It has to be done :) I went to BMAD repo to see how it goes, because today I decided to start that, BMAD harness for long run, I have few ideas, I need to read yours, thanks for that, and I need to clean mines. Look forward melting all that and see new insights from others . — Reply to this email directly, view it on GitHub [#1112 (comment)], or unsubscribe [https://github.com/notifications/unsubscribe-auth/ACPEXITYL5ALDYXQA3PDSOD4BNQZTAVCNFSM6AAAAACO4ADYOGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKMRUGQ3TQOA]. You are receiving this because you are subscribed to this thread. Message ID: ***@***.***>

[enving]

@coolcodemy please check this spam