Agile Agents

Overview

I worked on a multi-agent software engineering framework designed to explore whether structured collaboration between LLM agents can outperform single-agent systems on real development tasks. The project extended SWE-Agent with organizational coordination, role-based delegation, shared memory, and inter-agent communication inspired by human software engineering teams and Agile workflows.

At the center of the framework was an orchestration agent responsible for decomposing large software tasks into subtasks and coordinating specialized agents for planning, coding, reviewing, and testing. I helped build mechanisms for dynamic task assignment, centralized history management, summarization pipelines, and structured communication between agents operating over shared repositories and development environments.

One of the hardest engineering problems was maintaining coherent long-horizon execution. Agents would frequently lose context, enter edit loops, or fail to reason about how previous code modifications changed repository state. We introduced organizational memory systems, iterative summarization, and coordinated reporting pipelines to stabilize collaboration and improve shared situational awareness across agents.

We evaluated the system on SWE-bench using real GitHub issues and found that the multi-agent framework significantly improved true task resolution rates compared to single-agent baselines, despite lower self-reported completion confidence. The project made me much more interested in the future of agentic systems, especially around memory, coordination, planning reliability, and infrastructure for long-running autonomous workflows.