Agentic AI
TheBotCompany: LLM-powered agent teams tackle long-term software development
A new framework shows how self-organizing multi-agent systems can sustain continuous development over days, not just complete small tasks.
Researchers have unveiled TheBotCompany, an open-source orchestration framework that uses large language models to manage multi-agent teams through sustained software development cycles. Unlike prior work that treats coding as a series of isolated tasks, this system is designed to coordinate autonomous agents over extended periods—a gap that has largely gone unaddressed in the field.
How it works#
TheBotCompany operates through three core mechanisms:
- Three-phase state machine: Development moves through Strategy, Execution, and Verification phases, anchored to concrete milestones rather than ad-hoc tasks
- Self-organizing teams: Manager agents dynamically hire, assign work to, and retire worker agents as project demands shift
- Asynchronous human oversight: People remain in the loop without blocking agent progress at every step
Real-world testing#
The team evaluated the system on actual software projects over multiple days of continuous operation. Metrics included team adaptation patterns, milestone completion rates, cost efficiency, and code quality. The framework demonstrated that self-organizing agent teams can maintain productive long-term development and identified defects that a verification phase would have otherwise missed.
The work addresses a meaningful gap: while LLM-based systems have shown promise at automating individual coding tasks, persistent development has remained largely unexplored. TheBotCompany shows that with the right orchestration and oversight model, agent teams can sustain progress over time.
