The multi-agent trap every social team is about to fall into
Every social media operator I know is quietly running the same experiment right now: wiring a handful of AI agents into their content workflow and discovering that the bottleneck isn’t the agents — it’s the coordination between them. You ask one model to draft a TikTok hook, another to repurpose it for LinkedIn, a third to write the caption, and then you become the router, copying outputs between tabs and deciding which agent should handle the next step. That’s the real tax on AI-assisted publishing in 2025, and it’s the problem a small open-source project called Fez is trying to solve — not by making the agents smarter, but by making the room they work in smarter. If you publish across five platforms and juggle three or four AI tools, this launch is worth understanding, even if you never install it.
What Fez actually is, and the coordination problem it names
Fez is a chat-room-style workspace where AI agents are first-class members, each with its own identity, model, and skills, talking in the same channels you do. The maker is Ken, who frames the origin story bluntly: “I got tired of being the manager.” His complaint is one I’ve felt firsthand. Every agent app he tried either gave him a single assistant or gave him several and then made him do the coordination — pick an agent, repeat the question, read the result, decide if it was done, call the next one. “The agents were smart. The glue was me.”
That’s the sharpest line in the whole launch, and it’s the reason I’m writing about this at all. The “glue is me” problem is exactly what happens when a solo creator tries to build a content pipeline out of ChatGPT, Claude, and a couple of niche writing tools. The models are capable. The human is the message bus. Fez’s bet is that the message bus should be software.
The early version worked — agents in a shared room, each with its own identity — but the cost model broke. According to the maker, the only thing that could decide “who should take this?” was another expensive chat-model turn, so he kept doing it by hand. That’s a real architectural constraint, not a marketing complaint: routing decisions are cheap in human terms but expensive when you’re burning a frontier model’s tokens to answer a yes/no question. I’ve watched teams quietly rack up API bills doing exactly this — using a top-tier model as a glorified switchboard.
The judgment-model twist
The thing that changed the project, per the maker, was Jev, described as “a small judgment model from TypeSafe that answers yes or no with a calibrated probability, in under a second, for a fraction of a cent. It doesn’t write, it decides.”
That distinction — it doesn’t write, it decides — is the most interesting idea in the launch for anyone building content automation. It implies a split architecture: a cheap classifier handles routing, completion-checking, and reply-or-not decisions on every message, while expensive generative models only spin up when there’s real work to do. In the maker’s words, “the room makes three calls on every message: who takes it, is it done, does it need a reply. Chat models only run when there’s real work. I mostly just talk.”
Underneath, Fez runs on Nostr, the decentralized relay protocol. No account — your identity is a key, every agent has one, and everything lives on a relay you can run yourself. It’s MIT licensed and on GitHub, with a seven-minute tour in the media. Pricing, user counts, and roadmap are not disclosed in the launch.
How this differs from the tools you already pay for
Here’s where I have to be honest about the comparison landscape, because “AI agents in a chat room” sounds like a category that already exists. It mostly doesn’t — not in the form Fez is proposing.
The tools social teams actually use today fall into three buckets. First, scheduling and publishing suites like Buffer, Later, Hootsuite, and Metricool. These own the calendar and the API integrations to the platforms; their AI features are bolted-on caption generators, not autonomous decision-makers. Second, content production tools like Canva and CapCut, which own the asset layer. Third, general-purpose agent frameworks and chat assistants, which own the reasoning layer but have no concept of a shared workspace with persistent identities.
Fez sits in a fourth, thinner bucket: an agent coordination layer. It doesn’t schedule posts, it doesn’t edit video, and it doesn’t try to be your content strategy. What it does is give multiple specialized agents a shared room with identities, so the routing between them stops being your job.
My take: that’s a genuinely different value proposition from anything Buffer or Hootsuite is shipping, and it’s the layer most creator stacks are missing. But it’s also the layer where the incumbents have the least incentive to innovate, because coordination is invisible in a pricing page. Nobody upgrades a plan to buy “less glue.”
Why this matters more to a TikTok operator than a LinkedIn one
If you’re running a high-volume TikTok or Reels operation, your workflow has more decision points per asset than a LinkedIn ghostwriter’s does. A single short-form video spawns a hook variant, a caption, a hashtag set, a thumbnail concept, a cross-post to YouTube Shorts, and a repurposed carousel for Instagram. That’s six or seven routing decisions per post, and if you’re doing 30 posts a month across five platforms — a load I’ve run myself — you’re making two hundred micro-decisions a month about which tool handles what.
A LinkedIn-first operator publishing three times a week has maybe a tenth of that coordination surface. So the value of a judgment-model router scales with your publishing volume and your platform count, not with your follower count. That’s the operational lens I’d use to decide whether Fez is relevant to you: count your monthly routing decisions, not your audience size.
What creators and social teams can borrow from this, even without installing it
The most useful thing about this launch isn’t the product — it’s the architecture pattern. You can steal the thinking today, with the tools you already have.
Separate “decide” from “generate”
The single biggest cost leak in AI-assisted content workflows is using a frontier model for decisions that a tiny model could make. “Is this draft on-brand?” “Does this need a human review?” “Which platform should this go to first?” — these are classification problems, not generation problems. If you’re paying premium tokens to answer them, you’re overpaying. In my own tests of similar setups, the routing layer is where the bill quietly accumulates, because it runs on every item while generation only runs on the items that survive.
You don’t need Jev specifically to apply this. You can approximate it with cheaper models, keyword rules, or even a spreadsheet of if/then logic. The point is the separation of concerns: cheap decisions gate expensive work.
Give every agent a persistent identity and a narrow job
The “agents are members with their own identity, model and skills” framing is worth copying conceptually. When I’ve set up multi-tool content pipelines, the failures almost always came from agents with overlapping, vague responsibilities — two tools both “writing captions,” neither owning the final pass. Narrow roles with clear ownership reduce the coordination ambiguity that forces a human back into the loop.
Own your relay, own your workflow
The Nostr foundation is the part most social operators will skip past, and I think that’s a mistake to dismiss outright. Running on a self-hostable relay means your agent workspace isn’t locked to a vendor’s cloud, and your identity is a key rather than an account. For creators who’ve been burned by a tool shutting down mid-workflow — and if you’ve been in this industry more than three years, you have been — portability is a real feature, not an ideological one. Whether Nostr is the right substrate for a mainstream social team is a separate question I’ll get to below.
The repurposing angle
If you’re doing serious cross-platform repurposing, the room metaphor maps cleanly onto how a content team actually works: a writer agent, an editor agent, a platform-formatting agent, a scheduling agent, all in one channel, with a cheap router deciding who picks up each task. That’s a more honest model of the work than the “one magic assistant” framing every AI content tool sells. Here’s a related piece worth reading on how platform algorithms reward native formatting — the reason a single agent can’t just blast identical copy everywhere.
Where my judgment says this falls short
I want to be balanced here, because the launch is genuinely interesting and also genuinely early.
The Nostr bet is a double-edged sword. Decentralized identity and self-hosted relays are great for portability and privacy. They’re terrible for onboarding a social media manager who has fifteen minutes between client calls and needs something that works with the tools they already pay for. There’s no mention of native integrations with Buffer, Later, or any scheduling API, which means Fez is currently a coordination layer that sits next to your publishing stack, not inside it. That’s an extra tab, and extra tabs are where workflows go to die.
The judgment model is a single point of failure. A calibrated yes/no classifier is elegant until it’s wrong in a way you don’t notice — a draft marked “done” that shipped with a factual error, or a message routed to the wrong agent that stalls silently. The maker explicitly asks users to “run it and tell me where the room got it wrong,” which is refreshingly honest, but it also tells you the failure modes aren’t fully mapped yet. In my experience, silent misrouting is the hardest class of bug to catch in an automated pipeline, because nothing crashes — it just produces subtly wrong output.
It’s a developer tool wearing a creator-tool costume. MIT license, GitHub, self-hosted relay, Nostr keys — this is built by and for people comfortable in a terminal. The 7-minute tour suggests the maker knows onboarding is a hurdle. For a social media manager at an agency, the setup cost may exceed the coordination savings, at least in v1.
Pricing and scale are not disclosed. No user counts, no pricing, no roadmap. I’d bet the economics depend heavily on how cheap Jev actually is at volume — “a fraction of a cent” per call is great until you’re making three calls per message across a busy room. Where the math breaks is exactly at high message volume, which is… the scenario where coordination savings matter most.
Who this is NOT for
If you publish fewer than ten times a month on one or two platforms, you don’t have a coordination problem — you have a focus problem, and Fez won’t help. If you need a tool your whole team can adopt tomorrow with zero setup, this isn’t it yet. If you’re looking for content generation or scheduling, look elsewhere; Fez does neither.
What I’d watch / test next
The pattern here is more valuable than the product right now, so here’s what I’d actually do this week.
First, audit your own glue. For one week, log every time you manually move an output between two AI tools or decide which tool handles a task. If that number is under twenty, you don’t need agent coordination — you need to consolidate tools. If it’s over a hundred, you’re the exact operator this category is being built for.
Second, test the cheap-decision principle with what you have. Take one recurring routing decision in your workflow — “which platform does this asset go to first?” — and replace the model call with a rule or a cheap classifier. Measure whether output quality actually drops. I’d bet it doesn’t.
Third, watch the integration question. The thing that would make Fez or anything like it mainstream is a bridge to a scheduling API — Buffer’s API or Later’s — so the room can actually publish, not just decide. If that appears, the calculus changes.
Fourth, follow the judgment-model idea regardless of what happens to Fez. The split between “decide” and “generate” is going to show up in every serious content-automation stack within a year, because the economics are too obvious to ignore. Read the Fez launch page and the GitHub repo yourself, and if you’re technical enough to run it, take the maker up on his offer: run the room and tell him where it got it wrong. That’s how this category gets better.
