What is an agent fleet?
Published September 7, 2026 · by the AQ team
An agent fleet is a group of AI coding agents operated as a single unit: work is dispatched to them centrally, each agent runs in its own isolated environment, their state is visible in one place, and their output flows into one review process. The word marks a shift in how teams talk about coding agents. Through 2025 the unit of discussion was the agent (which one writes the best code); by 2026 the unit is the fleet (how many can you run, can you see them all, and who reviews the output). One agent is a tool you attend to. A fleet is a system you operate, and the vocabulary around it (fleet visibility, fleet handoff, fleet management) is operations vocabulary on purpose.
Why "fleet" is the word that stuck
Fleet is borrowed from operations, and the borrowing is precise. A vehicle fleet or a fleet of servers has three properties that a pile of individual units does not: the units are numerous enough that you manage them by policy rather than by hand, any single unit is replaceable, and the interesting state lives at the group level (utilization, failures, throughput), not inside any one unit. Cloud infrastructure normalized the usage years ago: AWS has offered EC2 Fleet and Spot Fleet APIs for launching and managing instances as one request since the late 2010s, and device fleet management is a settled category in IT.
Applied to coding agents, the same three properties hold. A developer running five agents in parallel does not read every token each one produces; they watch statuses and intervene on exceptions. A failed run is restarted or reassigned, not mourned. And the questions that matter are fleet questions: how many runs are active, which ones are blocked waiting for input, what merged this week, and what did it all cost. Teams reach for "fleet" exactly when those group-level questions start mattering more than any individual session, which is also roughly the point where a single developer stops scaling: how many coding agents one developer can run looks at where that ceiling sits and why.
Who says "fleet" as of September 2026
The term is now load-bearing marketing for several vendors, each attaching it to a different layer of the stack:
| Vendor | Fleet usage | What the fleet consists of |
|---|---|---|
| Orca | "The ADE for working with a fleet of parallel agents" | CLI coding agents in local or remote git worktrees |
| GitHub Agent HQ | Mission control to "choose from a fleet of agents" | Vendor coding agents run through GitHub |
| Warp Oz | "Orchestrating, observing, and scaling agent fleets" | Cloud agents across harnesses in Warp's platform |
| Cursor | Fleet management API for self-hosted worker pools | Worker machines that execute cloud agent sessions |
| LangSmith Fleet | Product named Fleet (renamed from Agent Builder, March 2026) | Enterprise AI agents generally, not only coding |
Orca, the open-source agentic development environment from Stably AI, is the loudest adopter: as of September 2026 its repository (roughly 63,000 GitHub stars) leads with "a fleet of parallel agents" and pitches fanning one prompt across five agents, each in its own isolated worktree, then merging the winner. GitHub announced Agent HQ at Universe on October 28, 2025, with mission control described as the place to choose from a fleet of agents (Anthropic's, OpenAI's, Google's, Cognition's, xAI's), assign them work in parallel, and track progress from any device. Warp launched Oz in February 2026 and as of September 2026 describes it as a platform for orchestrating, observing, and scaling agent fleets across harnesses and teams. Cursor's usage is the most literal infrastructure sense: its self-hosted cloud agent docs describe a fleet management API that autoscalers poll to size worker pools. And LangChain renamed its agent-building product to LangSmith Fleet in March 2026, aiming the word at enterprise agent operations broadly. The vocabulary converged before the products did, which is usually a sign a real category is forming underneath.
What operating a fleet actually requires
Strip the branding and a working agent fleet needs five things, whoever provides them:
- Isolation per agent. Two agents in one checkout corrupt each other's work. The standard mechanism is one git worktree and branch per run, with containers or separate VMs as heavier options. This is table stakes; every tool in the table above does some version of it.
- Dispatch. A way for work to enter the fleet: manually typed prompts, issues picked up from a tracker, or events that start runs automatically (see event-driven coding agents for the automatic end of that spectrum).
- Fleet visibility. One surface that answers "what is every agent doing right now": running, waiting on input, finished, or stuck, with the live output one click away. Visibility is the difference between operating a fleet and merely having started many agents; without it, parallel runs silently stall on permission prompts and you find out hours later. Monitoring multiple coding agent sessions covers the patterns in detail.
- Fleet handoff. Moving a live run between people and devices without losing it: the developer who started an overnight batch hands the fleet to a teammate, or picks the same sessions up from another machine the next morning. Handoff is what separates a fleet (an asset the team operates) from a set of personal sessions that die with one person's laptop. Handing off a coding agent session walks through what has to survive the transfer.
- Review at fleet scale. A fleet multiplies output, and every unit of output still needs a human decision before merge. This is where fleets quietly fail today: a July 2026 LeadDev analysis of 25,264 agent-generated pull requests across 2,361 popular GitHub repositories found that in 79 percent of agentic PRs the same developer both reviewed and modified the agent's contribution, and only about one in eight workflows involved multiple humans. Run a fleet without distributing the review load and one person becomes the merge bottleneck for the entire fleet's output.
Agent fleet, orchestration, and multi-agent systems
Three neighboring terms are worth keeping apart. Agent orchestration is the tooling layer that coordinates runs (isolation, queueing, tracking, review routing); the fleet is the thing being orchestrated. A multi-agent system is a research term for agents that communicate and negotiate with each other to solve one problem; a coding agent fleet is usually the opposite, many mostly independent agents on separate tasks that never talk to each other. And a fleet is not a swarm: swarm framing implies emergent coordination, while fleet framing implies central human control, which is the honest description of how teams run coding agents in 2026. The tool that operates a fleet made of different vendors' agents is what AQ calls a harness of harnesses.
Where AQ fits
AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. In fleet terms, AQ's position is that a fleet should be a team asset rather than a personal one. Agents run as real CLIs (Claude Code, Codex, Cursor Agent, Kimi, Grok, or plain shells) in persistent tmux sessions on your team's VM, each in its own isolated git worktree, so the fleet survives closed laptops and runs overnight. Fleet visibility is the dashboard: every workspace and its live session streamed to the browser, where any teammate can open the same workspace and watch the same live terminal. Fleet handoff is built in rather than bolted on: sessions resume from any device, and driving someone else's terminal is delegated by its owner approving a control request in one click. Dispatch comes from manual prompts or Linear (label an issue ai-task and a workspace appears, with status syncing both ways and humans starting agents, never auto-launch), and the review side is tracked per workspace as agents commit, push, and open PRs. If your fleet is one person's five local worktrees, a local tool is simpler; AQ earns its place when the fleet belongs to a team that needs to see it, share it, and hand it off. Start by running multiple AI coding agents in parallel.
Frequently asked questions
What is an agent fleet in software development?
An agent fleet is a group of AI coding agents a team operates as one unit: work dispatched centrally, each agent isolated in its own environment (typically a git worktree), all of them visible on one surface, and their output routed into a single review process. The term borrows deliberately from operations vocabulary like vehicle fleets and EC2 Fleet: many replaceable units, managed by policy, with the interesting state at the group level.
What is fleet visibility for coding agents?
Fleet visibility is one surface that shows the state of every agent at once: which runs are active, which are waiting on human input, which finished, and which are stuck, with each session's live output a click away. It matters because parallel agents fail quietly; a run blocked on a permission prompt looks identical to a run making progress unless something surfaces the difference.
What is fleet handoff?
Fleet handoff is transferring live agent sessions between people or devices without losing them: picking up your own overnight runs from a different machine, or handing a batch of in-flight sessions to a teammate at the end of your day. It requires sessions that live somewhere persistent (a server, not a laptop) and access control for who may view and who may drive each session.
How many agents make a fleet?
There is no threshold number; the honest boundary is behavioral. You have a fleet when you stop reading every agent's output as it streams and start managing by status and exception instead, which for most developers happens somewhere between three and five concurrent runs. Below that, ordinary terminal tabs work. Above it, you need the fleet layer: isolation, a status surface, and a review process that keeps up.
Is an agent fleet the same as a multi-agent system?
No. A multi-agent system is a research architecture where agents communicate and coordinate with each other on one problem. A coding agent fleet, as the market uses the term in 2026, is many mostly independent agents on separate tasks under central human control. The agents in a fleet typically never interact; the coordination happens in the tooling and the humans, not between the agents.