Run multiple coding agents at once: the limits by vendor

Once a background agent has fixed one ticket, the obvious next move is to hand it five. Whether you can run multiple coding agents at once, or "in parallel," depends on a number the pricing page often doesn't show: how many tasks the vendor lets run at the same time, and what it does with the sixth. Depending on the vendor, that extra task is refused, queued, billed, or started by pausing an older one. If you're sizing a plan, or you just hit a limit and want to know which one, here are the published numbers vendor by vendor.

We build Sinatra and sell plans priced on this number. Competitor numbers link to their own docs, and where a vendor publishes none, the table says "Not published."

Product details checked October 2026 against each vendor's own docs and pricing pages, linked throughout.

Agent	Tasks at once	Other caps	Past the limit
Google Jules	3 free, 15 on Google AI Pro, 60 on Ultra	15, 100 or 300 tasks per rolling 24 hours	New tasks disabled until the daily limit resets
Devin	Up to 10 on Pro and Max; unlimited on Teams and Enterprise	Daily and weekly usage allowance	Extra usage at API pricing
Kiro Web	10 tasks	Credits	Queued
Cursor Cloud Agents	No limit published: "as many agents as you want"	A spend limit you set	Not published
Claude Code on the web	Not published	Your plan's 5-hour and weekly limits	Usage credits, or wait for the reset
GitHub Copilot cloud agent	Not published	59 minutes per session, AI credits	A budget for additional usage
OpenAI Codex cloud	Not published	5-hour window on Plus and Business, weekly limits	Credits: bought on Plus and Pro, a workspace pool on Business
Linear coding sessions	Not published	Prepaid AI credit balance	Admins top up the balance
OpenHands Cloud	Not published	Not published	Older conversations paused, resumed when resources free up
Factory	Not published; depends on "organization compute limits"	5-hour, 7-day and 30-day rate limit windows	Rate limit error; missions pause
Sinatra	1 task at a time on Free, 3 on Pro, 10 on Team, 25 on Scale	No daily limit; 60 minutes per request on Free	Queued, up to 1, 10, 25 or 100; past that, refused with a message

Jules publishes the most. Its usage limits page lists "Concurrent Tasks" of 3, 15 and 60 next to daily tasks of 15, 100 and 300, measured over a rolling 24 hours. Past the daily limit "you will not be able to trigger new tasks until the limit resets," and the new task button is disabled. The page doesn't say separately what happens at the concurrent limit. Paid Jules plans come through Google AI Pro and Ultra, which Google says are "currently available only for individual Google Accounts (ending in @gmail.com)," so the 60 applies to one person's account. Our Jules review covers the rest of the product.

Devin's pricing page has a "Concurrent sessions limit" row: up to 10 on Pro ($20 a month) and on Max ($200), and unlimited on Teams and Enterprise. Free has no Devin Cloud at all. Usage is capped separately: Pro and Teams full seats have a daily and weekly allowance, Max a weekly one, and past that "you can purchase extra usage which is consumed at API pricing." Devin doesn't say what happens to an eleventh session on Pro.

Kiro Web runs 10 tasks at once. Its docs list a Queued state that "happens when you've reached the limit of 10 concurrent tasks," so the eleventh waits rather than failing.

Cursor puts no number on it: "You can run as many agents as you want in parallel." Spend is the limit instead. Cloud Agents use your plan's usage at the model's rate, and you set a spend limit the first time you use them.

Anthropic's answer for Claude Code on the web: "You can start multiple tasks and they'll all run simultaneously in separate sessions." Your Claude plan sets the limit. Anthropic says cloud sessions "share rate limits with all other Claude and Claude Code usage within your account," and "Running multiple tasks in parallel consumes more rate limits proportionately." Past the 5-hour or weekly limit you either turn on usage credits or wait for the reset.

Copilot caps each session at 59 minutes, a limit that "cannot be extended or bypassed," rather than capping the count: "You can run multiple agent sessions concurrently." Each task opens one pull request. Sessions draw AI credits, and past your included credits you "can continue working by setting a budget for additional usage," which GitHub notes "may be capped." Our Copilot pricing post has the credit math.

Codex cloud measures 5-hour and weekly windows. Plus and Standard Business have a 5-hour limit, Pro has none, weekly limits "may also apply," and "Cloud tasks may use more of your allowance than local messages." Plus and Pro users who reach the limit can buy credits, and on Business, work past a member's included usage can draw on the workspace's credit pool when flexible pricing is on, per OpenAI's Business credits article. We went through OpenAI's numbers in the Codex cloud pricing post.

For Linear, OpenHands and Factory we found no number. Linear coding sessions stop when the workspace's AI credits run out, until an admin tops up. OpenHands says that with "too many conversations running at once, older conversations will be paused," and a paused sandbox resumes automatically when resources are available. Factory says parallel launches "depend on organization compute limits," and past its rate limits "subsequent requests receive a Rate Limit error."

Jules, Factory and Linear refuse new work at their limits. Jules disables the new task button until the rolling 24 hours reset, Factory returns a rate limit error, and a Linear workspace with no credits left can't start a session until an admin tops up. In each case someone has to start the task again later.

Kiro and Sinatra queue it instead, and start the extra task when a running one finishes. Tickets assigned in a batch start on their own as room frees up, as long as the queue has space, and nobody has to retry them. Coder, which you host yourself, works the same way on its free license, with five agents at once: "Additional agents queue until an agent session becomes available."

Where the meter is usage, going past the included amount keeps everything running and adds to the invoice: Devin past its allowance, Copilot with a budget, Codex and Claude with credits, and Cursor up to the spend limit you set. You don't wait, and the bill changes from month to month with how much ran.

OpenHands pauses older conversations to make room for new ones and resumes them automatically when resources are available, so a long task can sit paused for a while because someone started another.

Sinatra's plans are priced on this number. Each plan is a flat price for the whole workspace, however many people are in it:

Plan	Price	Limit
Free	$0	1 task at a time, plus 1 queued
Pro	$20 a month	3 tasks at a time, plus 10 queued
Team	$99 a month	10 tasks at a time, plus 25 queued
Scale	$249 a month	25 tasks at a time, plus 100 queued

No plan has a daily or monthly task limit. Assign as many tickets as you want: the plan sets how many run at once, the rest are queued and start on their own as soon as one finishes, and when the queue is full Sinatra says so on the ticket instead of starting it. On Pro that message reads "Your workspace's queue is full (Pro plan: 3 tasks at a time, plus 10 queued), so I didn't start this one. Try again once a running task finishes." A queued task shows its position on the ticket. On Free, each request also runs for up to 60 minutes. If it needs longer, Sinatra stops, keeps the changes, and picks up when you reply.

A task counts as running from the moment Sinatra starts setting up its sandbox until the agent finishes that request. A task waiting for your reply, or sitting idle with a PR open, doesn't count, so ten PRs waiting for review use none of Pro's three. When you reply on one, it counts again, and it's queued if the plan is full. Requests from people start first, oldest first. Work Sinatra starts on its own, such as fixing a failed CI check, comes after them, and if the queue is full when you ask for something, your request takes the place of the most recently queued CI fix.

Tasks from Linear, from the Sinatra label on a GitHub issue, from Slack, and from the CLI, MCP and the API all share the one limit. The billing docs have the full rules.

Anthropic's sentence about parallel tasks applies well beyond Claude. If the model work runs on a subscription, five tasks at once spend that subscription's quota about five times as fast. On a plan with a 5-hour window, that can use up the window well before it resets.

That's true on Sinatra too. Each member's tasks run on their own Claude, ChatGPT or SuperGrok subscription, or on a workspace Anthropic or OpenRouter key, and Sinatra never bills for model usage. Three tasks from one engineer on Pro share that engineer's subscription limits. Spreading work across the people who started it spreads it across their subscriptions, and a shared API key has no 5-hour window at all, just a per-token bill from your provider.

Size for how many tasks you'd want running at the same moment on a normal busy day, rather than how many you'll hand over in a month. On a small team that might be one or two for each person who delegates, and tasks waiting for review don't count toward it.

Then look at review. Every task that finishes is a pull request somebody has to read. If two people review, ten tasks at once mostly produces a longer review queue, and reviewing AI-generated PRs is still the slow part.

On Sinatra, a team with two or three people delegating fits Pro: 3 tasks at a time, plus 10 queued. A queue that's always full is the signal to move to Team. If you're also weighing seats against usage pricing, the per-seat vs per-workspace post works through a five-person team.

Jules Ultra runs 60 tasks at once for one person, the highest published number here, and its free tier runs 3. Devin Teams has no limit on sessions. Cursor, Claude Code on the web and Copilot publish no cap on the number of tasks, so what you can run is bounded by what you're willing to spend or your plan's rate limits.

Sinatra's Free plan runs 1 task at a time, against 3 on Jules' free tier. On any Sinatra plan, work past the limit is queued, and there's no option to pay per task to run it sooner. If you'd rather pay to finish a busy week faster, a usage-priced agent lets you.

The pricing page has every plan. The free one needs no card, so the quickest way to learn how much your team runs at once is to start free and assign it a few tickets.