The Missing Primitive for Agent Swarms — Lou Bichard, Ona
Summary
Lou Bichard from Ona discussed the "missing primitive" for agent swarms in the context of building software factories, which aim to incrementally remove human intervention from the SDLC. While runtimes, orchestration, and triggers for coding agents are largely solved, the critical challenge lies in agent coordination. Ona's platform supports agent fleets, enabling automated tasks like CVE remediation across thousands of repositories by spinning up isolated development environments (VMs) for sub-agents. Key difficulties include breaking down the SDLC into micro-steps, managing context rot in LLMs, and the inadequacy of existing human-centric tools like GitHub for agent coordination. Proposed solutions for coordination include state machines, durable executions, and CLI constructs.
Key takeaway
For AI Engineers and MLOps teams building autonomous software factories, your primary focus should shift from agent runtime and orchestration to developing robust agent coordination layers. Relying on human-centric tools like GitHub for inter-agent communication creates overwhelming noise and inefficiency. Instead, investigate implementing state machine workflows or CLI-based coordination protocols to manage complex agent interactions, ensuring deterministic behavior and effective context management across the SDLC.
Key insights
Effective agent coordination, not runtime or orchestration, is the missing primitive for fully autonomous software factories.
Principles
- Software factories automate SDLC by incrementally removing human interaction.
- Harness engineering embeds process knowledge into repositories for agent guidance.
- VMs provide essential security and performance isolation for agent runtimes.
Method
Implement agent fleets by defining prompts and scripts to automate tasks across multiple repositories, utilizing parent agents to spawn and manage sub-agents in isolated VMs for parallel execution and message passing.
In practice
- Automate large-scale code changes like CVE fixes using agent fleets.
- Deconstruct complex SDLC stages into granular micro-steps for agent training.
- Explore state machine or CLI-based approaches for inter-agent coordination.
Topics
- Agent Swarms
- Software Factories
- Agent Coordination
- SDLC Automation
- Ona Platform
- Harness Engineering
- Virtual Machines
Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, MLOps Engineer, AI Architect
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.