Walk through collapsing an aesthetic two-agent split back into a single-agent flow

Short answer·Hard·4.0 · 0·~3 min·Asked atAnthropic OpenAI Cursor

Attempt it

A team built a 'research agent' that calls a 'summary agent' for every customer-support response. The split has no measured quality win over a one-agent baseline; it just felt cleaner to write. Walk through the migration back to a single agent, what you preserve, and what you delete.

Free · 2 AI evals / day

**Step 1: Prove the split is aesthetic.** Build a single-agent variant with one system prompt that has a research section and a summary section. Give it the same tools as the research agent. Run the existing eval set on both the old two-agent system and the new single-agent variant. Compare quality, cost, and latency. **Step 2: Decide what to preserve.** Keep the prompts (they become sections of one larger system prompt), keep the tool list, keep the eval set, keep the cost and latency dashboards. The content is good; only the wiring is wrong. **Step 3: Decide what to delete.** Delete the handoff tool, any supervisor or routing logic, and the second agent's separate observability surface. The per-agent step-correctness scores are now meaningless because there is no second agent. **Step 4: Roll out carefully.** Do not just flip the switch. Run the single-agent variant on shadow traffic, canary at 5 percent, ramp to 100. Keep the old multi-agent path warm for one release in case the eval missed something. **Step 5: Tell the truth.** 'The split was wrong, here is the data, the simpler thing performs as well.' That is good engineering, not failure.

**Start with the proof, not the surgery.** Build a single-agent variant whose system prompt has two clearly named sections (a research phase, then a summary phase) and the same tool set as the existing research agent. Run the **existing eval set** on both systems. If the single-agent variant ties or beats end to end quality at lower cost and lower latency, the split is confirmed aesthetic and the collapse is safe. **Preserve what was actually working.** The prompts are still good content; they become two sections of one system prompt. The tool subset, the eval set, the cost and latency dashboards all carry over unchanged. None of these were the problem. **Delete what only the multi-agent shape required.** The handoff tool, any supervisor or routing logic, the second agent's separate observability surface, the per-agent step-correctness traces (now meaningless without a second agent to attribute to). These were wiring, not value. **Roll the collapse out like a forward migration in reverse.** Shadow traffic first, canary at 5 percent, ramp to 100 over a few days. Keep the old multi-agent path warm behind a kill-switch flag for one release in case the eval missed a long-tail case. **Tell the truth to the team.** 'The split sounded clean. The data says it did not earn its keep. Here is the simpler thing that performs as well.' Rolling back complexity that did not pay is engineering maturity, not a confession.

Quick reference

TL;DR

Prove the collapse with an eval, preserve prompts and tools as sections of one system prompt, delete the wiring, and ship the rollback with the same canary discipline as a forward migration.

Memory aid

Easy to grasp

Imagine an office where every email goes through a researcher who writes a draft and then a summarizer who shortens it. Someone notices that one well-trained person writes the same quality email faster and cheaper. The right move is not to fire the summarizer in a panic. It is to give one person a checklist with two clearly named steps (research first, then summarize), prove the new setup writes equally good emails on the same test inbox, keep the two-person setup warm in case the test missed something, and roll out the change carefully. The story the team tells is honest: the two-person split sounded clean but did not add value. Removing it is not a failure; it is the proof working.

Key concepts

Multi-Agent Systems

Detailed answer & concept explanation

~5 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Quick scan · 2–4 min

Everything important, quickly.

7 min: prove with an eval first + preserve prompts and tools as sections + delete wiring and per-agent traces + shadow canary ramp the collapse + write an honest postmortem that encourages future cleanups.

Real-world examples

Real products, models, and research that use this idea.

Anthropic's research and summary patterns in production deployments often start as single-agent flows precisely to avoid the aesthetic-split trap.
LangGraph and CrewAI both let you express 'two phases in one agent' via prompt sections and tool subsets, which is exactly the collapse target.
Sierra and Intercom Fin have publicly described iterating from multi-agent designs back to leaner single-agent ones when evals showed no quality win.
Internal eval harnesses at scale-up AI teams routinely include 'single-agent ablation' as a comparison baseline so an aesthetic split gets caught early, not after launch.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QWhat signals would tell you mid-rollout that the eval missed something?

QHow do you write a postmortem that encourages future collapses instead of discouraging them?

QWhen is a measured-zero quality win still worth keeping the split?

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Ripping the second agent out without an eval run first, so when quality regresses on a long-tail case there is no proof the collapse was safe and no warm fallback to flip back to.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

References

Primary sources. Skim if you want the original framing.