liveBook · Manning

Multi-agent system

As explained in Vibe Engineering

A multi-agent system is a software system in which multiple autonomous or semi-autonomous agents operate concurrently or collaboratively toward specified objectives.

A multi-agent system coordinates multiple agents toward specified objectives. In AI-assisted development, coding agents can work in parallel on different parts of a codebase while a developer reviews and merges their output. Parallelism offers leverage, but isolated agents may repeat context-gathering and produce conflicting changes. section 9.5.3

Use parallel agents where the work has clear boundaries:

  • Split tasks by responsibility and, where possible, by non-overlapping files—for example, an API endpoint, a frontend component, and integration tests. section 9.5.3
  • Isolate workspaces, typically with Git worktrees, and budget for repeated reads of shared code. When agents edit the same utility, conflict resolution and reruns can erase the benefit of parallelism. section 9.5.3
  • Prefer a single agent working sequentially when tasks are tightly coupled or boundaries are fuzzy. Evaluate orchestration tools only when enough independent work exists to keep several agents busy. section 9.5.3

The cost is not just sessions and developer review. Multi-agent workflows commonly use four to seven times the tokens of a single-agent session and can introduce stuck subagents, corrupted plan files, or looping verification. Missing context can trigger expensive retries. For a simple bug fix or quick refactoring, that overhead can be excessive. section 3.5.3

Ambiguity at scale

In a web-browser experiment, hundreds of concurrent agents produced over three million lines of Rust in a week. The risk is that unclear instructions and implicit conventions scale too: agents can generate plausible, passing code with subtle architectural inconsistencies that compound across commits. Make constraints and conventions explicit before multiplying agents, and retain human judgment over their design. section 10.8.1

Measure the coordination cost rather than assuming parallel work is cheaper. One team found that roughly 20% of two-agent sessions ended in conflicts requiring manual resolution, often costing more than sequential execution. section 9.5.3

How Vibe Engineering covers it

How deep it goes
Explains task decomposition, workspace isolation, cost tradeoffs, and specification risks, without providing an implementation walkthrough.
Examples
Uses an API/frontend/test split, a measured two-agent conflict rate, and a concurrent Rust browser-building experiment.
What you can do after
Evaluate multi-agent coding work against sequential execution using task independence, coordination costs, and specification clarity.
How it connects
Connects multi-agent work to Git worktrees, merge conflicts, context engineering, and architectural drift.

Also explained in

Compare how 13 books explain multi-agent system →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Vibe Engineering. The section links point to the parts of the book it draws on.