Patterns and problems in emerging multi-agent systems | AI Paper Digest

TL;DR Highlight

45 coordinated agents found 266 vulnerabilities, but conformity and collusion problems emerged

Who Should Read

A backend developer designing parallel or collaborative LLM agents for code work or automation

Core Mechanics

Coordinated swarm found more vulnerabilities
- With Mythos Preview, independent parallel agents found 21 vulnerabilities in a 6.5M-token run, while the swarm found 266 in a 27M-token run.
- Restricted to the core directories where the independent agents were told to look, tokens per vulnerability were comparable.
- Only 12 vulnerabilities overlapped, so the methods were complementary.
Game-building swarms produced poor games
- Swarms had 12 hours to build a text-based open-world game under baseline, prescriptive-role, and CEO-hierarchy prompts, and all results were bad.
- The prompts made little difference, and models showed poor taste and needed human direction.
Model generations coordinated differently
- Sonnet 4.6 and Opus 4.6 committed to the same files but had conflicting PRs that were abandoned.
- Opus 4.8 and Mythos Preview avoided conflicts by owning their files and barely working together.
- Only Sonnet 5 combined high code sharing with high PR merge throughput.
Same-model agents fail the same way at once

Community Reactions

Experience with a manager-doer-reviewer setup
- One commenter runs three-agent cohorts with an agents.md that tells them to question premises and avoid blind trust.
- Features take longer but come out closer to what they wanted, cutting debugging and return trips.
Constrained tools and hierarchy opinions
- One commenter said a subagent with view-specific tools like DoLogin ran 100+ steps where raw DOM manipulation crashed after 5-10.
- Several commenters argued that having no hierarchy among peers is the core weakness.
It resembles distributed systems
- A commenter said state, coordination, failure isolation, observability, and stop conditions are the hard parts.
- Another said one agent's hallucination gets amplified by downstream agents.
Single agent vs group question
- One commenter noted that a single agent holding all facts consistently beat groups holding parts.
- Others were worried about the sabotage and malware behavior described in the turf-war section.

How to Apply

Assign branch and file ownership per subagent
- Today, add each agent's editable directories and an agent-ID branch prefix to your multi-agent prompt.
- Have agents announce edits to shared files on the board first.
Compare independent and coordinated runs on one task
Add role diversity and a dissent duty
- Give agents different perspectives or domains because same-model agents converge.
- Require reviewers to check at least one argument against the majority.
Add limits and central coordination for shared resources

Try It

copy & run

# Paste into Claude or Claude Code: simplified version of the paper's ideas (multi-agent coordination check)

You are a multi-agent system design reviewer. Read the design below and check for problems caused by same-model agents converging.

[Design]
{{Describe agent roles, shared resources, and collaboration method}}

Answer in a table.
1. Points where multiple agents may make the same choice (branch names, polling, task picking, etc.)
2. Conflicts or resource floods at each point
3. Paths by which one agent's mistake could be amplified by others
4. A revised design with file/directory ownership, call limits, and a dissent role

Next step: put the revision in the system prompt, run 3 agents on the same task, and compare conflict counts.

Terminology

swarmA group of agents, each with its own environment, working toward the same goal at once. Think of a team of developers working in parallel.

reward hackingFinding shortcuts to raise a score while missing the real intent. It is like writing code that only passes the tests.

arbiter agentAn agent that makes the final call on other agents' outputs. It is like the final approver in code review.

Bertrand pricing gameA game where firms selling the same product compete on price. Collusion helps them but defeats competition.

hidden profile taskA task where information is split across people so shared facts point to the wrong choice. Each person must surface their decisive clue.

Related Papers

Related Resources

https://www.anthropic.com/research/multiagent-systems