Devin Bots and Replit Missions both shipped yesterday. They hand an agent team a goal for weeks, and the controls you set first matter more than the model

Cognition and Replit launched long-running agent teams on November 17, 2026. Devin Bots own an ongoing engineering responsibility inside your repo and Slack. Replit Missions chase a stated outcome on a fixed budget. Before you try either, decide what counts as done, what needs your approval, and the spending cap.

On this page
  1. What changes from a normal agent session
  2. Side by side
  3. The controls that matter
  4. What the launch numbers show, and don't
  5. What it costs to find out
  6. Before you start one
  7. Which one fits
  8. Sources

Cognition's launch post describes a Bot as a persistent agent with its own memory and tools that can lead several Devins in parallel. Replit's Missions post describes a goal you define, which Replit then works toward by coordinating a swarm of agents to plan, build, test, and iterate. Both are a step past the single-task agent session most teams use today.

They aren't alone. Fortune's launch story frames Devin Bots as a rival to OpenAI's Dots, which works with Codex, and Anthropic's Duet, which works with Claude Code. We haven't run Bots or Missions. Everything below comes from the vendors' posts and docs, and the numbers are theirs.

What changes from a normal agent session

A normal session takes a task, does it, and stops. Both products replace the task with something that stays open:

  • Devin Bots take a responsibility. Cognition's examples turn “Port this feature to Android” into “Keep Android at parity with iOS”. The Bot files tickets, starts a Devin session for each one, checks the results, and decides what to do next.
  • Missions take an outcome. Replit's example is “win our first 10 paying enterprise customers”. Setup asks questions to pin down what counts. An independent AI judge checks results against those criteria, and the Mission ends when they're met or the budget runs out.

That second example matters. Missions isn't only for code. Replit's agents research buyers and draft outreach alongside building a demo. Devin Bots stay inside software engineering.

Side by side

Devin BotsReplit Missions
Unit of workAn ongoing responsibilityAn outcome with success criteria
How it endsIt doesn't; the Bot keeps runningCriteria met, or budget spent
Who checks the workYour CI, Devin Review, and the evaluation loop you set upAn independent AI judge against the agreed criteria
Where it talks to youSlack, iMessage, the Devin app, and a “Needs You” inboxUpdates you choose, sent immediately, daily, or weekly
MemoryRetains context across sessionsAgents share findings within the Mission
Spending controlBilled to the Bot's creator; dollar or ACU caps on Teams and EnterpriseA spending limit set in the Missions agreement
PlansMax ($200/month) and Teams ($80/month plus $40 per full seat)Not stated in the launch post

Sources for the table are Cognition's launch post, its pricing page and Bot usage limits docs, and Replit's Missions post.

The controls that matter

Both vendors put the same three decisions on you. They are where a long-running agent goes right or wrong.

What counts as done. Cognition tells engineers to define success criteria for a Bot's scope, such as which CI checks must stay green, and to build the feedback loop from tests, benchmarks, user flows, and production behavior. Replit's setup interview does the same job. Vague criteria are the main way either product burns money on the wrong thing.

What needs approval. Replit's example sets a $1,000 cap and requires approval before the agents contact prospects, share a proposal, change customer-facing materials, or make product commitments. Devin Bots let you set which systems a Bot can reach, what it can change, and when spending or releases need sign-off. Per Fortune, teams can also auto-approve low-risk edits.

Where it runs. Fortune reports that Bots run on separate cloud machines, away from credentials on engineers' laptops, with admins choosing which repos and tools each Bot can use. It also reports network controls meant to block data leaving if a model follows injected instructions. Replit's agents also run in the cloud, but its post doesn't describe network controls.

What the launch numbers show, and don't

Cognition's headline case is an Android Bot started on September 28, 2026. After 44 days the app was in open testing on Google Play. Cognition reports 1,134 merged PRs, 344,651 lines of Kotlin, 2,841 screenshot tests, and 1,217 closed tickets, with the engineer sending 183 messages in total.

Read the merge rule closely. The Bot merges PRs that pass CI and Devin Review, so most of those 1,134 PRs had no human reviewer. That works if your tests are strong. If they aren't, a Bot that merges dozens of PRs a day turns weak test coverage into a lot of unreviewed code. Bilt says its Bots merged about 1,600 PRs in a week; ask what your team would do with that volume.

Replit's headline case is DataCat, an internal monitoring platform built in two days on a $3,000 budget. Replit compares it with a build from the same prompt in Claude Code using /goal:

MetricMissions (DataCat)Claude Code build
API compatibility tests passed51 of 5645 of 56
Trace intake49,000/sec9,800/sec
Metric intake405,100 points/sec224,600 points/sec
Storage enginesPostgres and ClickHouseSQLite

Replit is upfront about the limits: one run per setup, scored only on shared product areas. It also says the Claude Code build covered more product areas and ingested logs faster. One run each is an anecdote, not a benchmark. Replit's NumPy test is more systematic, with a 32-agent swarm reaching up to 19x speedups against 4x for one agent across 13 tasks.

What it costs to find out

Devin Bots bill to the person who created them, including the Bot's chat and every session it starts, according to the usage limits docs. On Max, usage comes out of your included quota, then on-demand credits at API rates. On Teams, a creator or admin can set a dollar on-demand cap, but it only limits on-demand spend: at the cap, the Bot keeps running on the creator's included quota. For a real stop, an admin can give the Bot a separate usage limit, which moves all its usage to on-demand and stops it at the cap. On Enterprise, a Bot stops at its ACU or dollar limit, but only if budgets and user limits are enabled for the enterprise.

That last condition is the one to check. An Enterprise Bot without budgets and user limits enabled has no spending stop. Replit's design is simpler: the budget is part of the Mission agreement, and work stops when it runs out, whether or not the goal was reached. Replit's Missions docs add that work already running can take spending slightly over the budget before it stops. Replit's post doesn't say which plans include Missions or how usage is priced.

Before you start one

  1. Pick a scope you'd give a contractor. “Keep the Android app at parity” or “cut p95 page load below 2 seconds” works. “Improve the product” doesn't.
  2. Write the success test first. Name the CI checks, benchmarks, or metrics the agent must hit. If you can't measure it, neither can the judge.
  3. Set a cap that actually stops work. On Replit, set the Mission budget a little under what you'd write off, since running work can overshoot it. On Devin Teams, have an admin set a separate usage limit for the Bot; the on-demand cap alone doesn't stop it. On Enterprise, confirm budgets and user limits are enabled before relying on a Bot limit.
  4. List the actions that need you. Releases, migrations, anything customer-facing, and anything touching auth or billing.
  5. Decide who reviews merges. If the agent can merge on green CI, check that your tests would catch the mistakes you care about.

Which one fits

Pick Devin Bots if you have an existing codebase, solid CI, and a recurring engineering job nobody wants to own, such as platform parity, performance regressions, or dependency upkeep. It fits teams that already live in Slack and GitHub.

Pick Replit Missions if the goal is a measurable outcome you can state upfront and fund once, especially greenfield builds or work that mixes code with research and outreach. The fixed budget makes it easier to try without a long commitment.

For teams already on Codex or Claude Code, it may make sense to look at Dots or Duet before switching vendors. We'll compare all four once we have hands-on results.

Sources

Researched and drafted by an AI agent from the linked sources. How posts are made.