The takeaway
Multi-agent orchestration is now a shipped product, and the binding constraint has moved from model quality to cost per accepted result.
Why it matters for builders
Model tiering is now an architecture decision, not a cost heuristic: pin the strongest model to coordination and verification, run cheap models on parallel execution, isolate every agent's state, and gate deterministic steps outside the agent loop. Multi-agent runs multiply token spend by an order of magnitude, so cost per accepted change is the metric to optimise.
Google's Agent Swarm Moves Multi-Agent AI From Demo to Product
Google made Gemini 4 Carbon available to every paying developer and enterprise today, roughly two weeks after handing it to a small group of cybersecurity defenders through its Fairwind program. The model is the headline in most timelines. The more consequential thing shipped alongside it: /swarm, a preview feature in Google Antigravity that turns a single prompt into a coordinated team of agents with a live board, a roster of pinned models, and dedicated adversarial verifiers.
That is Google productising multi-agent orchestration, and it makes the economics of agent work impossible to ignore.
What Google Actually Shipped
Carbon is positioned as the model for work that "tests the limits of a model's intelligence" - multi-step coding and debugging, open-ended scientific research, cybersecurity. Google keeps Gemini 4 Argon in the lineup and says plainly that Argon remains the right pick for most software engineering and knowledge work. Carbon is the expensive, thorough option: it explores longer, reasons deeper, and runs extended verification loops, which means it burns more tokens per task.
Google claims first place on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, at more than two times lower cost per task than the next best external model. Pricing for the API is $2 per million input tokens and $10 per million output tokens as an introductory rate that doubles after November.
That price path is the tell: Google is buying adoption for three weeks and betting teams are locked in by then.

How /swarm Is Actually Built
The Antigravity team's own write-up is unusually specific, and the engineering choices are the interesting part.
/swarm sends your prompt to a family of agents rather than a chain. A coordinator spins up specialized parallel workers in isolated git worktrees so their edits cannot collide. Adversarial verifiers are spun up separately and their job is to run tests and challenge diffs, not to help write them. Every agent posts to a shared, append-only message board, and idle agents park at zero token cost, reactivating the moment another agent needs their input.
That last detail is the one builders should steal. Idle agents that still hold context and still bill are the hidden cost centre in most homemade multi-agent systems. Parking them at zero and waking them on demand is what makes a wide swarm affordable at all.
Agents can be pinned to different models. Google recommends Carbon for coordinator, architecture, security-audit and verification roles while lighter, faster models take the parallel file edits and routine implementation. That is model routing expressed as an org chart, not as a cost heuristic bolted on afterwards.
The harness is also deliberately visible. Antigravity ships a Swarm side pane with a live Board feed, a Members roster you can jump into to steer an individual agent, and an interactive Graph timeline showing how the team converged from brief to verified code. The CLI gets a full-screen board via /swarm board or Alt+S. Google's showcase demo - a 906,000-triangle Three.js reconstruction of an office built from seven planning PDFs across ten overnight loops - is less persuasive than the fact that the run can be inspected while it happens.
Google says its engineers spent five weeks dogfooding swarms on migrations, root-cause investigations and architectural redesigns.

The Coordination Tax Comes Due
Google's own documentation includes a warning that deserves to be quoted in every architecture review: even with zero-token idle waits and dynamic model hot-swapping, parallel multi-agent systems consume quota much faster than a standard chat.
That single sentence reframes the whole agent conversation. For two years the cost story was price per million tokens. Swarms break that metric. A five-agent team where the coordinator is on the expensive model and verifiers re-run the full suite can spend ten to fifty times the tokens of a single agent on the same task, and the only number that matters becomes cost per accepted change.
Which is why the first benchmark Google chose to lead with is cost per task, not tokens. It is also why the pricing doubles at the end of the month. Multi-agent work is a volume business and volume businesses are priced after the habit forms.
There is a second cost that does not appear on any invoice. Orchestration quality now dominates model quality. A frontier model in a badly coordinated harness loses to a mid-tier model in a good one, because the failure modes are coordination failures: two workers editing the same file, a verifier that approves its own team's work, a coordinator that summarises away the detail the next agent needed. Google's answers - isolated worktrees, adversarial verification separate from authorship, an append-only shared log - are structural rather than prompt-level mitigations.
The append-only board has a side effect for regulated buyers: a replayable record of which agent claimed what and when.

What Builders Should Take From This
Most n8n Lab readers are not going to run Antigravity swarms. The patterns transfer directly to workflow automation, and they are cheaper to adopt than the model upgrade.
Pin models per role. Planning and final review on the strongest model you can afford, extraction, routing, and formatting on the cheapest model that passes your eval. Wide agent teams amplify a wrong routing decision across every branch.
Isolate state, not just prompts. The git worktree decision is a general principle: parallel agents must not write to the same resource. In automation terms that means separate records, separate branches, separate queues, and an explicit merge step that a human or a verifier approves.
Separate verification from authorship. A checker agent from the same context that produced the output will ratify its own mistakes. Verification needs a different prompt, ideally a different model, and read-only access to the artifact.
Make the run observable. A shared log every agent appends to, with timestamps and actor names, is worth more than any dashboard. It is also the only way to debug a swarm after the fact.
Gate the expensive paths deterministically. Deterministic steps are free and repeatable; agent steps are neither. Every place where an if-statement would do is a place where a swarm should not be spending tokens.
Access is limited. /swarm is a preview inside Antigravity for Google AI Ultra subscribers, with enterprise access through Gemini for business, and there is no public API for the swarm harness itself. Teams building on n8n, LangGraph or their own orchestrator are reproducing these patterns by hand, and today's release is Google's reference implementation for how to do it.
What to Watch
Three things decide whether this lands. First, independent cost-per-task data once the introductory pricing ends this month. Second, whether /swarm escapes Antigravity into the Gemini API or Google Cloud, because an orchestration layer that only exists inside one IDE is a feature, not infrastructure. Third, competitive pressure: Anthropic shipped Fable 5.2 a day earlier and reportedly leads on general intelligence indices, so the next round will be fought on harness quality and price per accepted result rather than raw model scores.
The practical read: the model is good, the price is temporary, and the orchestration patterns are the durable part.
The Automation Brief
Read 5 AI stories instead of 50.
The essential moves in AI agents, models, automation and infrastructure — filtered for builders and operators, with the part that actually matters.
No noise. Unsubscribe anytime.
Editorial notes
Stefan Trbojevic
n8n Lab Editorial
10 November 2026
10 November 2026
Sources
AI disclosure: AI assisted with research and drafting. Factual claims are reviewed by an editor.




