Skip to main content
Back

When to Add a Second AI Coding Agent (Reliable Multi-agents)

Two Hacker News posts say a team of Qwen3.8-27B models matches or beats Claude Fable 5. Here are the four reasons a second agent helps, and how to test yours.

  • In the GVS5H paper, a team of Qwen3.8-27B models scored 92.4 versus 66.8 for a single call on 100 hard LiveCodeBench problems, but its 92.4 to 90.4 edge over Claude Fable 5 carried a p-value of 0.88, far from statistically significant.
  • memdoor's independent rebuild of the GVS5H loop cost a median of $0.39 and 40 minutes of model time per problem on OpenRouter, with the longest run taking 4 hours.
  • With thinking off, one call of Qwen3.8-27B solved 35 of 100 problems while its multi-agent workflow solved 47, gaining 17 problems but losing 5 the single call had gotten right.
  • Anthropic's Haiku 5.5 update on 4 November cut status-line update costs by about 99%, translating to roughly 5% less total usage in sessions with background subagents.
  • An Amazon-internship study found a single-conversation agent slightly outperformed a multi-agent team on HumanEval (92.1 vs 91.6) while costing less ($0.020 vs $0.026), and a matched-thinking-budget test showed single agents beating sequential teams 0.418 to 0.379 at 1,000 tokens, though teams won at a 100-token budget (0.364 vs 0.290).

Recently, Hacker News put the same claim on its front page twice: a team of copies of Qwen3.8-27B, a small open model with 27 billion parameters, matches or beats Anthropic's Claude Fable 5 on hard coding problems.

First came the GVS5H paper on 25 October, then memdoor, an independent rebuild, on 31 October.

And under both posts sits the question every developer with a coding agent now has: when does a second agent actually help?

An agent, here, is a model running in a loop with tools: it reads files, runs commands, and edits code until the job is done. Its context window is everything it can see at once, it's like a working memory for the job.

And a team, or multi-agent setup, gives several copies of a model one job each (plan, write, review) and passes the work between them through files.

So my answer, from the papers, the docs, and the setups that ship code every day, is that a second agent helps only when it gives you something one agent can't have.

There are four such things: a separate context window for a side task, independent tasks that can run at the same time, a different model, or, just as important, a review by an agent that didn't write the code.

A second agent that gives you none of those is one agent's work at a higher cost. So for every second agent you run, the decision is Keep it, Remove it, or Test it first.

And on 22 October, Anthropic's Applied AI team wrote that across several enterprise deployments, one agent with skills (instruction files it loads when needed) "frequently outperformed" the subagent design, where a main agent hands jobs to helper agents.

What follows is what a second agent costs, the four reasons with a case for and against each, the two Hacker News results as the run files show them, and a 10-task test with a checklist to copy.

What does a second AI agent cost?

A second agent costs tokens, time, and some detail at every handoff, all paid before it returns anything.

As you know, a token is the unit of text a model reads and writes, and the unit you pay for. Each Claude Code subagent "sends its own requests, which count toward the same usage limits," per the subagents docs, and GVS5H's manager-and-worker loop raised output tokens "to about 3x baseline."

Even watching a subagent costs tokens. In Claude Code's terminal app, a running background subagent updates its status line every 30 seconds, and until 4 November each update re-read the subagent's whole conversation.

That day, Anthropic's ClaudeDevs account said Haiku 5.5 now writes the line from short notes, so each update "costs about 99% less." In sessions with background subagents, that came to "about 5% less of the limit," it added.

Then comes the second cost: time. memdoor's rebuild of the full GVS5H loop took a median of 40 minutes of model time and $0.39 per problem on OpenRouter with the longest run at 4 hours.

Last comes the handoff: one agent passing work to another as a summary, which is most of what the next agent knows. Anthropic's guide says agents split by problem type "engage in a 'telephone game,' passing information back and forth with each handoff degrading fidelity."

Its one measured figure is about speed: one customer saw "a roughly 2x faster runtime" from a single agent loop than from its multi-agent harness, the setup of prompts, tools, and loop around a model.

So say you're about to add a reviewer agent to your Claude Code setup: after the main agent writes a change, a second agent reads it and says whether it's right. Every section from here asks whether that reviewer earns its cost.

When should you add a second agent? Four reasons

Add a second agent when it serves one of four reasons:

  • a separate context window for a side task
  • independent tasks that can run at the same time
  • a different model
  • an independent review

Each gets a case where it helped and one where it didn't, with numbers, so you can tell which case is yours.

1. A separate context window for a side task

A separate window pays for itself when a side task would bury the main conversation. Claude Code's rule: use a subagent "when a side task would flood your main conversation with search results, logs, or file contents you won't reference again," and keep only its summary.

And Lauren Tan (@poteto), who works on Grok Bot and wrote the pstack plugin for Cursor, put that rule in its instructions: "Route bulk to subagents, keep summaries in the main thread."

Anthropic's guide agrees. Its first case for keeping a subagent is "measurable performance degradation from context bloat," and its third is an agent that struggles to pick the right tool or carries conflicting instructions.

But a separate window stops paying when every "agent" is the same model with a different prompt, because one conversation can then play all the roles.

A paper from an Amazon internship, on the older GPT-4o-mini, ran the same multi-step workflow two ways: as a team of agents, and as one agent taking each role in turn in one conversation.

On HumanEval, a benchmark of short Python functions, the single conversation scored 92.1 to the team's 91.6. By the paper's cost estimate, it also ran for $0.020 instead of $0.026.

That is an older small model on short problems, so the direction is the result to keep.

So your reviewer should read the change in a fresh window, with nothing from the main agent's conversation. That's pstack's "Fresh subagents by default" in one line.

2. Independent tasks that can run at the same time