Rethinking AI Coding: Multi‑Agent Collaboration as the New Development Paradigm

Why Single‑Agent AI Coding Falls Short

Traditional AI IDEs treat the human and a single agent as a linear partnership, where the agent receives a request, modifies files, runs commands, and returns results. This creates a bottleneck as developers spend most of their time waiting for the agent’s output and manually reviewing changes.

Adopting a Multi‑Agent Collaboration Model

To improve efficiency, the author experiments with multiple concurrent agents, assigning each a distinct task, reviewing their work, and integrating results. This approach demands tooling that supports parallel agents, centralized review, and visibility into each agent’s activity.

Evaluating Existing Tools

OpenCode’s web mode offers a glimpse of the desired workflow by allowing agents to run for a period before the user intervenes for feedback and acceptance. However, it lacks true parallelism, prompting the need for a custom solution.

Introducing Mexus: A Web‑UI for Managing Parallel Agents

Mexus is designed not as another AI IDE but as a web interface that lets a single user orchestrate multiple agents simultaneously. It can run locally or on a server and presents a dashboard‑style layout with three panes: a left pane listing CLI agents, a central pane showing agent activity and code diffs, and a right pane displaying the workspace file tree.

Spec‑Driven Orchestration

Instead of dispatching tasks directly, Mexus first generates a structured spec that defines goals, scope, acceptance criteria, impact, and suggested task breakdown. After user review and approval, an execution plan is created, assigning each agent specific responsibilities, allowed file paths, and dependency relationships, with two mandatory human approval checkpoints.

Soft Boundaries and Claims

Each agent receives an allowedPaths list restricting its file modifications. During execution, a lightweight claim records which agent is editing which files. Claims are not hard locks but provide real‑time visibility of file usage, helping detect potential conflicts.

Observer Agent for Runtime Coordination

An additional Observer Agent monitors the environment, detecting overlapping edits, out‑of‑bounds changes, or stalled tasks. It can automatically suggest avoidance actions for minor conflicts or alert the user for critical issues, operating under strict rate limits and audit logging.

Review as the Central Workflow

In Mexus, review is continuous rather than a final step. The central pane displays diffs from each agent, allowing developers to approve, reject, or send feedback directly back to the corresponding pane, keeping the loop within the review interface.

Managing Shared Workspaces vs. Git Worktrees

While Mexus supports git worktrees, the default is a shared workspace to encourage collaboration. Shared workspaces risk concurrent edits, but Mexus mitigates this with allowed paths, claims, and the observer, ensuring real‑time conflict visibility and resolution.

Why Build a New Tool?

Existing solutions either focus on terminal session management or generic agent consoles. Mexus aims to place review at the core, provide a unified UI for task distribution, execution monitoring, and result acceptance, and leverage web‑based interactions for lower entry barriers and richer UX.

Implications for Human‑AI Collaboration

The shift toward multi‑agent systems moves human effort from writing code to designing environments, defining intents, and steering agents. Structured constraints (allowed paths, claims) become amplifiers that ensure reliable, scalable agent behavior, aligning with the “Harness Engineering” concept of humans steering and agents executing.