An Accidental Blackboard
Thoughtworks engineers, experimenting with 'hyper-agentic engineering,' accidentally stumbled upon an effective coordination mechanism for AI agents working in a shared monorepo. This emergent behavior, reminiscent of a classic 'blackboard system,' allowed agents to self-organize and build a complex airline IROps system in just four days. The story offers a fascinating glimpse into the future of AI-driven software development and collaborative agent intelligence.
The Lowdown
The article details an experiment by Thoughtworks Europe where a team of 10 engineers, using 'hyper-agentic engineering' with AI agents, built a complex airline IROps (Irregular Operations) system in just four days. The most significant outcome was the unexpected discovery of an emergent coordination pattern among the agents.
- The team's objective was to rapidly develop a highly complex airline IROps system, which manages disruptions like technical faults or sick crew members across hundreds of aircraft and thousands of passengers.
- They employed a monorepo setup, with all agents working concurrently, and introduced a strict discipline of continuous commit and rebase to main.
- This discipline unexpectedly led agents to use the shared plans and code in the monorepo as a 'blackboard system' to track each other's progress and coordinate interdependent tasks.
- Agents would update plans to reflect progress, allowing others to see when dependencies were being met or when work areas were claimed, facilitating an organic integration process.
- The author recognized this as an accidental re-discovery of the 'blackboard system' or 'tuple space' pattern, a well-known architecture for coordinating autonomous problem solvers.
- While effective, this accidental solution had drawbacks, such as overwhelming the CI pipeline with frequent commits, and was not an intentionally designed system.
- Inspired by this emergent behavior, the author is now developing 'Talwrn,' an intentional blackboard system designed specifically to provide a structured communication channel for agentic engineering.
This experiment underscores the potential for emergent intelligence in agentic systems and highlights the importance of shared context for sophisticated AI collaboration, even if initially unintended.