Multiple sessions or multiple agents — which actually speeds you up?

These two are often treated as two ways of doing the same thing. In fact they solve problems along two different axes — they don't conflict, and neither can replace the other.

One-line answer

Answer Our takeIt's not either-or: multiple agents (subagents) solve width within a single task — side quests like searching, testing, and research run in parallel in separate contexts, bringing only summaries back to the main conversation; multiple sessions solve length across time — each independent piece of work gets its own nameable, resumable session. Ask "is this a side branch of my current task, or a separate piece of work?" and the answer falls out.

Steps

Classify the work first.OfficialThe criterion for a subagent: a side task would flood the main conversation with search results, logs, or file contents you'll never reference again — so it does the work in its own context window and returns only a summary. Conversely, if it's a separate piece of work unrelated to what you're doing now, it deserves its own session.
For width within one task, use subagents.Just say it in the conversation: "use subagents to research the auth, database, and API modules in parallel".OfficialSince v2.1.198 subagents run in the background by default, so the main conversation isn't blocked; but note that each subagent's result eventually flows back into the main conversation — spawn too many, each returning a long report, and they crowd out your context all the same.
For multiple tasks across time, use multiple sessions.OfficialA session is a conversation bound to a project directory and persisted to disk: name it at startup with claude -n auth-refactor, then resume it by name with claude --resume auth-refactor; inside a session, /branch copies the current conversation to try another path while leaving the original untouched. For the full session-management playbook, see D6 · Session management.
Multiple sessions editing the same repo will step on each other's files — the fix is one worktree per session; see Worktrees: keep parallel sessions out of each other's way.

At larger scale, upgrade by asking "who coordinates".OfficialThe docs give selection criteria for four parallelism modes:

Mode	Who coordinates	When to use it
Subagents	Claude dispatches and collects results within one conversation	Side tasks would flood the main conversation
Agent view (`claude agents`)	You toss independent tasks into the background and check status later	Several unrelated things in flight at once (research preview)
Agent teams	A lead agent splits work, assigns it, and tracks progress	You want Claude itself to coordinate a group of workers (experimental, off by default)
Dynamic workflows	A script holds the plan and runs dozens or hundreds of subagents	Repo-wide audits, large migrations — work beyond what one conversation can coordinate

Our own experiment backs this up.Our takeWe ran a controlled experiment with 27 agents and reached the same conclusion: multiple agents manage width within one task, multiple sessions manage length across time and total context, and the two axes are orthogonal; what actually decides whether parallel development succeeds is having an interface contract everyone follows. Full write-up: Multiple sessions or multiple agents? I ran 27 agents to find out.
Cost note.OfficialRunning several sessions or subagents at once multiplies token consumption, and it all counts toward your plan's usage and rate limits.

Copy-paste prompt

I have these things on my plate:
1. <task A>
2. <task B>
3. <task C>

Classify them first — don't start working yet:
- Which are side branches of the same task that can run in parallel?
  Handle those yourself with background subagents; each subagent should
  return only its conclusions and a list of changes, not its working
  output;
- Which are independent tasks? List those back to me and I'll open a
  separate session for each with claude -n <name>.

Show me the classification for confirmation first; once I confirm,
execute only the parts that belong in this session.

Sources & last verified

OfficialManage sessions, fetched 2026-08-05.
OfficialCreate custom subagents, fetched 2026-08-05.
OfficialRun agents in parallel, fetched 2026-08-05.
Our takeMultiple sessions or multiple agents? I ran 27 agents to find out, our long-form experiment write-up, verified 2026-08-05.
Last verified: 2026-08-05 · volatility:high (agent view and agent teams are both preview/experimental features; behavior changes across versions).