Why Qoder gives every AI agent a boss

When multiple AI agents try to work together on a complex engineering task, the natural instinct is to let them talk directly. Share findings, request help, broadcast updates. But as the number of agents grows, that web of conversations quickly becomes a mess: conflicting information, redundant work, a state space that balloons out of control, as explored in analysis of parallel agent orchestration.

Alibaba Cloud's Qoder takes the opposite approach. Its new Experts Mode replaces peer-to-peer messaging with a single agent that acts as Team Lead. The Leader decides who does what, when, and stitches the results together. The agents never talk to each other. They talk only to the boss, a design that mirrors Microsoft's thesis on orchestration over raw model power.

One brain, many hands

Experts Mode splits the work into a DAG of dependent tasks. The Leader maintains a global view, breaks requirements into subtasks, and dispatches them to specialized experts, a structure that avoids the planning traps that trip up less regimented agent systems. These experts run asynchronously and in parallel: once the Leader sends a task, it immediately moves on to the next one, so multiple experts can work simultaneously.

The five built-in expert types each come with their own toolset:

• Research Expert, technology selection, root cause analysis, competitive research, codebase impact assessment. It integrates semantic search, code search, knowledge graphs, and web search. Qoder claims a 21% F-Score improvement over industry baselines for code search.
• Coding Expert, new features, emergency bug fixes, refactoring, API development, database modeling. Runs in a cross-platform sandbox (Mac, Linux, Windows) for security.
• QA Expert (Verify), change-aware regression verification, quality gates, functional validation, pre-deployment checks, addressing the reality that AI generated code is twice as likely to contain bugs.
• Code Review Expert, PR/MR review, risk assessment, security and performance checks, architecture review. According to Qoder's internal benchmarks, it ranks Top 2 in the industry, beating Cursor and Codex.
• Browser Expert, frontend interaction testing, end-to-end form flows, visual regression, cross-browser checks. Qoder claims it saves 60% in token costs compared to Playwright's MCP-based approach, a significant saving given the opaque cost structures of many AI coding agents.

How coordination works

Each expert reports back through a mailbox. After finishing a task, an expert sends a callback that lands in the Leader's inbox. The Leader, when idle, reads the message, runs model reasoning, and decides the next action: dispatch another expert, update the task state, or reply to the user. This loop (report → evaluate → act) continues until all tasks are done.

The user stays in the loop too. You can interrupt at any time, and the Leader will incorporate your input in its next cycle, adjusting direction or canceling work. The role shifts from driving every step to aligning intent and reviewing outcomes, more like leading a real engineering team.

Learning from each task

After completion, the system detects signals (corrections, failures, user instructions) and extracts reusable skills. Both the primary agent and the sub-agents capture domain-specific skills from their own work. These skills are stored as persistent memory and recalled automatically when similar tasks appear, forming a closed learning loop.

Benchmarks: quality and cost

Qoder tested Experts Mode on an internal benchmark of complex engineering tasks covering full-stack development, cross-language refactoring, and feature implementation. They evaluated across four quality dimensions: usability, completeness, UI experience, and engineering quality. Experts Mode outperformed on all four. The gains come from the architectural choices, centralized coordination, async parallel scheduling, and specialized experts, rather than just adding more agents.

The data, of course, is Qoder's own, and the comparison model is unspecified. But the principle is defensible: in systems where agents have overlapping capabilities, a shared leader can maintain coherence that free-for-all messaging struggles to preserve. The architecture mirrors a real engineering org chart for a reason.