Last week a reel crossed my feed: a dashboard with some 300 AI agents, each in its own window, all working at once. The creator called it a fleet. It looked like the future, and I've been thinking about it ever since — because I believe the demo showed me the wrong half.
The industry is optimizing for the spawn. Bot fleets, group chats full of agents, "just spin up another one." Every demo stages the same moment: look how many we can run in parallel. The implied thesis is seductive — agent power equals agent count.
I'm betting against it. Spawning is the easy 10%. The real product — the part nobody demos — is convergence: taking N conflicting, overlapping, differently-formatted agent outputs and collapsing them into one human's decision context. Collect, connect, organize, unify. Without convergence, a 300-agent fleet is a very expensive screensaver.
What I learned in the wrong room
I spent years in telecom standards meetings (3GPP, CT3 group). The hard part of cellular was never the radio. It was getting base stations to agree — handovers, synchronization, consensus procedures. Connectivity without coordination is just interference: everyone transmitting, nobody heard.
Agent swarms today are at the "everyone transmitting" stage. The demos celebrate parallel work the way early telecom celebrated parallel radios — before the industry understood that the product was agreement, not transmission.
I'm now building that agreement layer myself, inside a personal AI system designed for exactly one user. I recently started on the output-convergence piece: merging multiple agent outputs so they land in one person's context with provenance intact. And that constraint — one user — taught me something the fleet demos never will: convergence needs a destination. One inbox, one decision, one person accountable for what happens next. A swarm with no convergence target isn't a fleet. It's noise multiplied.
The refutable part
Disagree with me — please. The obvious objection: convergence is just summarization, and we already have that.
No. Summarization compresses text. Convergence resolves conflicts. It ranks evidence by trustworthiness, keeps provenance (which agent said what, when, with what source), and knows what to escalate versus what to discard. Summarization is a text operation; convergence is a decision operation. One makes the noise shorter. The other makes it actionable.
The next real milestone in agent infrastructure won't be announced as "we can now run 1,000 agents." It will be announced as "ten agents, one answer." Whoever ships that ships the product. The rest of us are just spawning.
Jiahui Miao is building vertciti, starting from a personal AI system designed to serve exactly one person. He writes about what the build teaches him.
Top comments (4)
The telecom comparison lands on the right distinction. When multi-agent systems try to solve convergence by passing prose into another prompt, the agreement layer just becomes an expensive summarizer that hallucinates away edge cases. The only way I keep subagent runs tractable is treating convergence as a protocol problem using typed receipts with explicit artifact hashes and exit reasons. If two workers disagree on facts, the system flags a merge conflict in the ledger rather than asking a judge model to split the difference.
Everyone might be broader term here. There are two streams of agentic development going on.
These are pure tech companies - they have resources and pretty much at mature level (mostly these companies with research teams) - these guys know how to establish the co-ordination.
Other companies - they just want to use tags and the work force here has "self agenda" - so this is the where your point is pretty much visible.
Four agents, all producing valid outputs, and the merge step failed every time because two of them had tagged the same company under different names with no way to detect it. We spent about a week on prompt changes before someone noticed the keys didn't match across agent outputs at all. Adding a pre-merge normalization pass that resolved names to canonical IDs before any aggregation ran fixed it, but it took a surprisingly long time to notice that was where the problem was sitting.
The convergence problem also seems to require an explicit definition of what “agreement” means. Two agents can produce different outputs that are both internally valid, while two seemingly similar outputs can hide a meaningful disagreement in assumptions or evidence. I’d make the merge layer responsible for identifying those differences rather than forcing everything into a single answer. That could mean comparing structured claims, source provenance, confidence, and unresolved assumptions before aggregation. The important part is that convergence shouldn’t always end in consensus—sometimes the correct output is a clearly surfaced conflict that gives the human enough context to decide. That makes the convergence layer a decision-support boundary rather than just another processing step.