Short version: between July and August 2026, the three coding agents most solo builders live in each shipped a way to run work in parallel — and they made three different bets. Codex multi-agent V2 is a fleet you command. Claude Code turned separate sessions into a mesh that can talk. Cursor side chats are a quiet scratchpad that reads and answers without touching your build. They look like the same feature. They are not. Pick by control model, not feature count — and watch the token bill, because every one of these multiplies spend as fast as it multiplies throughput.

Codex multi-agent V2: the fleet with the most knobs#

In August, OpenAI stabilized the opt-in multi-agent V2 experience in Codex (OpenAI changelog). The headline is granularity: each spawned sub-agent thread can carry its own model, reasoning level, and role, bounded by a per-session concurrency cap — what one analysis called per-thread runtime routing. You can put a frontier model on planning and a cheap one on mechanical edits, in the same run.

That power is also the catch. The V2 rollout drew explicit developer concerns about transparency — it's easy to fan out work, and spend, that you can't easily see. Codex now warns when high concurrency plus a high reasoning setting could run up usage fast. Reach for it when you're parallelizing one large task and genuinely want the routing knobs; skip it if you won't watch the meter. This is the same mixed-model discipline we mapped in how to reduce AI agent token costs and why cheap models fail silently in long agent loops.

Claude Code: a mesh, not a tree#

Claude Code already had subagents — workers spawned inside one session (and their cost math is real). What changed on August 7 (v2.1.224) is coordination between sessions: with ListAgents a session can discover your other running Claude Code sessions, and with SendMessage it can deliver a message to one of them, across machines on macOS and Linux (9to5Mac; releases).

That's a different topology. Codex gives you a directed tree — a root that commands sub-agents. Claude Code now gives you a peer mesh — long-lived sessions (say, one per service) that hand off to each other as equals. If your work is already spread across several terminals, this is the reason to stay; we broke down the mechanics in cross-session messaging with SendMessage and ListAgents, and the subagent-control tradeoff in Claude Code vs Cursor vs Cline.

Cursor side chats: the safe one, on purpose#

Cursor's move is the opposite of a swarm. In 3.11 (July 10), typing /side or /btw opens a side chat: a durable, full conversation that inherits your main chat's context but, by default, only reads, searches, and answers (Cursor changelog). You can revisit it later and @-mention it to pull its findings back into the main thread.

The point isn't to scale the work — it's to scale your attention. When you hit "wait, how does this API actually behave?" mid-build, a side chat lets you chase the tangent without stopping or corrupting the agent that's shipping. Read-only-by-default is the whole design: it can't quietly refactor something while you're not looking.

So which do you reach for?#

Subagents and fleets scale the work. Side chats scale you. Buying the wrong one is how a "productivity feature" turns into a bill you can't explain.

The meta-point survives the next release on all three changelogs: parallelism is a cost decision before it's a speed decision. Turn on cost attribution first, decide who's actually in control, and only then fan out. If you're still choosing your primary terminal agent, start with Cursor vs Windsurf vs Copilot vs Claude Code and Devin vs Codex vs Cursor vs Jules on background agents — then come back here to decide how it runs in parallel.