Five agents in a group chat is not an architecture. It is a meeting that bills by the token.
The first thing most teams do after they get one agent working is add a second. That is usually the mistake.
Last Friday I wrote that Claude Code sessions can now message each other on the same machine, and that people will try to turn that primitive into a team. I have already watched the next move in Slack threads and architecture reviews. Researcher agent. Writer agent. Critic agent. Editor agent. Manager agent. They sit in a group chat and talk until a termination condition fires. The transcript looks like collaboration. The invoice looks like a standing meeting that never learned to adjourn.
The instinct is human, not stupid. Work scales by adding specialists and putting someone in charge of the handoffs. Multi-agent marketing maps onto that picture so cleanly that it feels like architecture. A prior-authorization packet splits into eligibility, clinical criteria, coding, and a determination letter. A claims file moves through intake, adjudication, and correspondence. The org chart is sitting right there. The framework will even let you name the agents after the desks.
Communication is the part of the mapping that does not survive contact with a model. People share a room, a case file, and a sense of when talking has stopped helping. Agents share a context window. Every turn one agent spends “discussing” with another is a full prompt of someone else’s conversation, billed twice, with no shared instinct to shut up. I have sat through enough incident reviews to know what cheap talk looks like in a human team. Agent-to-agent chatter is the expensive version of the same failure.
The demo hides this because the demo is a single happy path with a narrator. Five agents debate a pull request. The critic finds a bug. The writer patches it. The manager declares victory. Nobody measures the tokens that went into the debate, the latency of the extra hops, or the number of times the critic restated a finding the writer already had in its system prompt. Production does not come with a narrator. Production comes with a budget, an SLA, and an auditor who wants to know which component made the call.
That is the week’s question. Not whether multi-agent systems can look impressive. Whether the coordination pattern matches work that actually splits.
There are four patterns worth taking seriously, and only some of them earn their keep.
Supervisor is the one that works. One agent routes. Specialists execute. The supervisor is a dispatcher, not a debate moderator. It receives a job, decides which specialist owns the next interface, and collects a structured result. Eligibility in, clinical criteria out. Criteria in, determination letter out. Each specialist has a job, a schema, and a stop condition. The supervisor does not ask them to discuss. I have used this shape on workflows that already had human desks in the same topology. The agent version works for the same reason the human version works: the work was already decomposable, and the interfaces were already named.
Hierarchical is supervisor with extra floors. A manager delegates to workers, who delegate to sub-workers. It works when the work actually nests. A FOIA request that fans out across three departments, each of which has its own intake rules, is a tree. A single clinical judgment is not. Every extra level costs you a hop of latency and a hop of context. The manager summarizes for the director. The director never sees the finding, only the summary. That is how humans lose the plot in a large organization, and it is how hierarchical agents lose the plot faster, because the summary is generated, not written by someone who will get paged if it is wrong. Use the tree when the work is a tree. Do not use it because an org chart is comforting.
Swarm is the pattern I keep getting pitched and keep declining. Agents self-organize. They pick their own next speaker. They form and dissolve coalitions. Research papers love this because it photographs well and produces long transcripts. Production hates it because there is no contract. You cannot tell an auditor which agent decided what. You cannot tell a debugger why the loop ran seventeen turns instead of four. You cannot write a test for a termination condition that lives in the emergent behavior of the group. Outside a lab with a human watching every turn, swarm is a group chat with extra vocabulary. I will cover the cases where a constrained swarm is not theater later this week. Those cases are narrower than the marketing.
Blackboard is the pattern most teams skip, and it is the one I would start with after supervisor. Agents do not talk to each other. They read and write a shared workspace. The workspace is a schema, not a transcript. Agent A posts a structured finding. Agent B consumes the finding, not Agent A’s chain of thought. The claim file is a blackboard. The incident channel with a pinned status is a blackboard. The ticket that carries state between desks is a blackboard. Humans already coordinate this way when the work is serious. We do not hold a five-person standup to update a field on a case. We update the field. The next person reads the field.
The reason blackboard gets underrated is that it looks less like AI. There is no witty cross-talk. There is no screenshot of agents complimenting each other. There is a record. That record is the product. If you need an audit trail, you already wanted a blackboard and you built a chat instead.
A blackboard also fails in a way you can see. If the schema is wrong, the next agent refuses the write or writes garbage you can point at. If a group chat is wrong, the error is a vibe in turn fourteen. I would rather debug a rejected JSON document than a transcript in which three specialists agreed with a hallucination because agreement is what group chats optimize for.
The decision rule is blunt. Ask whether the work splits into independent subtasks with named interfaces. If you can name the interface without using the word “discuss,” supervisor or blackboard will carry it. If the subtasks nest, hierarchical will carry it, and you should budget the extra latency in advance. If the only interface you can name is a conversation, you want one agent with better tools and a better prompt. Adding agents will not create a decomposition that the work does not have. It will create tokens.
I keep seeing the same failure in architecture reviews. The team cannot get a single agent to produce a reliable determination letter, so they add a critic agent to review the letter, then a manager agent to reconcile the critic and the writer. Three agents now share one unsolved problem. The letter is still unreliable. The bill is not. The move that would have helped was specifying the letter, not staffing it. A schema, a checklist, a retrieval set that actually contains the policy language. Multi-agent coordination cannot compensate for a missing interface. It can only hide the absence behind a transcript.
The cost shows up before the accuracy problem does. Two specialists plus a supervisor is three model calls where one used to live, plus whatever they write back into the next prompt. Give each specialist the prior transcript “for context” and you have reinvented a meeting that reprints the minutes at every seat. I have watched teams celebrate a better-looking answer while the p99 on a workflow that used to take one hop doubled, then tripled, because nobody put a budget on talk. Latency is the tell. If the second agent did not remove work from the first, it added delay.
There is a version of this that is not theater, and it is why the week exists. Microsoft folded AutoGen and Semantic Kernel into the Agent Framework. The Python package hit 1.14.0 yesterday. Magentic, Handoff, and GroupChat now sit in the box as named orchestration patterns rather than forum folklore. AutoGen’s own repo still has the larger community, about 60,400 stars, but its last Python tag is 0.7.5 from last September and the code has not been pushed since April. The merged product is where the commits are going. That stack is the enterprise bet: auth, telemetry, middleware, C# and Python in the same SDK. It is also the stack whose group-chat inheritance I want to interrogate, not applaud. Shipping GroupChat as a first-class pattern makes theater easier to adopt. Handoff is closer to supervisor. Magentic is the one I will want numbers on before I trust it with a case file.
CrewAI is the other honest evolution. The project is at 1.15.16 as of yesterday, about 57,100 stars, and it still ships the role-based crews that made the demos easy. Flows put a deterministic path under those roles when the path is the thing compliance cares about. Role-play for the work that is actually creative. A flow for the work an auditor will ask you to replay. That hybrid is the right concession for a framework that started as a casting call.
Thursday is DSPy. Version 3.3.0 left beta on August 3. Instead of hand-tuning how agents talk to each other, you declare the output you want and the compiler searches the coordination. The learning curve is still steep and the community is still smaller than the chat-first frameworks. The instinct is correct. If the talk is the expensive part, you should not be writing it by hand.
Between those tool posts I am going to argue the other side out loud. Multi-agent is the wrong answer more often than the timeline admits. A single well-tooled agent beats a committee on cost, latency, and debuggability for most of the work people currently assign to committees. Wednesday I will put the pattern that survives that argument into a reference architecture for a regulated workflow: supervisor, sandboxed specialists, a validation gate, durable execution, and an audit log that would make a CISO less unhappy than a group-chat transcript.
The hinge from last week still holds. Once sessions can talk, teams will build teams. Some of those teams will match work that already had interfaces. Most of them will be five agents burning context to rediscover a checklist. The pattern that works is the one that reduces conversation. The pattern that does not is the one that adds agents until the demo looks like a department.
If you are about to stand up a second agent, name the interface first. If you cannot, you do not have a coordination problem. You have one agent that is not finished.
If this was useful, forward it to one engineer who needs less noise in their feed.