Replit Missions: Power-User Guide to Agent Swarms, Budgets and the AI Judge

Last updated: November 20, 2026

Replit Missions, launched November 17, 2026, lets you hand Replit an outcome instead of a task. You agree on success criteria, a dollar budget and approval rules in a Mission contract, then a swarm of cloud agents plans, builds, tests and iterates until an AI judge confirms the goal is met or the budget runs out.

Contents

What Replit launched

Replit announced Missions in an official blog post on November 17, 2026, updated November 19. You define the outcome, and Replit coordinates a swarm of agents to plan, build, test and iterate toward it within the budget and boundaries you set.

Unlike a normal Agent request, a mission keeps its goal attached to the chat across turns and sessions, and it can pause and resume while it waits for your input.

Missions extends Replit’s recent push into autonomous agents. For background, see our guide to Replit cloud agents and our Replit Agentic Commerce and Ads guide, which covers the November 12 purchasing and advertising features.

Missions vs Routines vs tasks

Replit now has three ways to delegate work. This table summarizes Replit’s Missions vs. Routines vs. Tasks documentation:

Feature Best for What you define How you review results
Missions Achieving a goal with agreed success criteria Goal, approach, success criteria, budget and boundaries Review the deliverable and the Mission report
Routines Repeating work on a schedule Instructions and a recurring schedule Review the result of each scheduled run
Tasks Making a specific project change The change you want and its plan Review and apply the task’s changes

How to start a mission

These steps follow the Missions documentation:

  1. Open a chat and state the goal. Type /mission followed by your goal, ask Replit to start a mission, or click New in the sidebar and select Mission.
  2. Choose a planning depth. Work through the details asks more questions; Only the essentials moves faster to the agreement.
  3. Understand. Explain what you want, why it matters and what you are starting with, so Replit can scope the work.
  4. Approach. Pick one of the suggested approaches or describe your own.
  5. Success. Define outcomes that can be checked against the result, not just activities to perform.
  6. Boundaries. Set the maximum Replit AI usage budget in USD, select the model, and list actions that need your approval or must never happen.
  7. Review and start. Check the Mission contract in the side panel, ask for revisions, then select Confirm and start. A contract marked Draft is not active.

Writing a strong Mission contract

Replit’s guidance and examples point to a few patterns:

  • Make success measurable. “Pages respond in under 400 ms and checkout tests pass” is checkable. “Improve performance” is not.
  • Name the evidence. Ask for test results, benchmark output or a comparison against a baseline so the judge has something concrete to check.
  • Fence off risky actions. Replit’s sales example requires approval before the agent contacts prospects, shares a proposal, changes customer-facing materials or makes product commitments.
  • Say what not to guess. Replit’s launch-plan tutorial tells the agent to identify missing information rather than guess, and not to contact anyone, make purchases or publish anything.

A sample goal in that style: /mission Get every page on our marketing site to load in under 500 ms on the agreed test setup. Keep all existing tests passing. Ask before changing anything customer-facing. Do not deploy.

You can change the agreement later. Ask Replit to revise the goal, success requirements, permissions or budget, then review the updated contract. According to the Missions product page, you can also choose which connectors a mission may access and exit mission mode at any time.

Budget, spending and pricing

Replit has not published a per-mission price. Its product page FAQ says you set a maximum budget upfront and that work performed during the mission counts toward it. Messages you send while a mission is active also count.

The product page shows spending by step and says any increase to the limit needs your approval. The documentation adds that running work can push spending slightly over budget, and the budget does not guarantee success.

Replit’s blog examples give a sense of scale: a $1,000 cap for an enterprise sales mission, and a $3,000 budget for the two-day internal build described below. For how other vendors meter multi-agent work, compare our Perplexity Computer Multiagent guide.

Progress updates and the AI judge

During planning you choose which updates you want, including blockers, progress updates and general status, and how often: immediately, in a daily digest or weekly. Per the launch post, the cloud agents keep working after you close your laptop.

When the agents try to finish, an independent AI judge checks the evidence against your success criteria. It either accepts completion or sends back specific feedback, and the agents revise. Per the product page FAQ, the mission keeps working until the criteria are met, its budget runs out, or it needs your input.

Clearing a mission is not the same as stopping it. The documentation warns that clearing the goal does not cancel background processes or reverse actions already taken, so explicitly ask Replit to identify and stop ongoing work if needed.

What Replit says swarms achieved

Replit published several results. These are vendor-run tests, so treat them as indicative:

  • NumPy optimization: across 13 benchmark tasks, a single agent on Auto (Max) reached speedups of up to 4x, a 16-agent swarm 17x and a 32-agent swarm 19x. On Auto (Power) the figures were 2x, 7x and 9x.
  • DataCat: in two days, within a $3,000 budget, Missions built an application monitoring platform with dashboards, alerts and tracing. It flagged a planted database bug in 38 seconds against a 60-second target and passed 51 of 51 data and alert checks.

Replit also compared DataCat with a platform built from the same prompt using /goal in Claude Code. DataCat accepted logs, metrics, host data and traces, while the Claude Code build accepted logs only; the Claude Code build covered more product areas. Replit notes this used one run per setup. For the closest Anthropic feature set, see our Claude Code Projects guide.

Limitations to plan for

  • No published pricing or plan list: budget in Replit AI usage dollars, and confirm costs before large missions.
  • Automated completion is not proof: Replit says the judge’s verdict does not guarantee the result is correct; review it against your criteria.
  • Unfinished outcomes: if the budget runs out first, work stops with the goal unresolved.
  • One-off by default: missions do not repeat; use a Routine for recurring work.

Frequently asked questions

What is Replit Missions?

Replit Missions, launched on November 17, 2026, gives a Replit chat a persistent goal. You agree on an approach, success criteria, a USD budget and boundaries, and a swarm of agents plans, builds, tests and iterates until an AI judge confirms the criteria are met or the budget runs out.

How do I start a mission in Replit?

Type /mission followed by your goal in a chat, or ask Replit to start a mission. You can also click New in the sidebar and select Mission. Replit then walks you through planning before you confirm the Mission contract.

How much does Replit Missions cost?

Replit has not published a separate Missions price. You set a maximum Replit AI usage budget in US dollars for each mission, and all work, including messages you send while the mission is active, counts toward it.

How does Replit decide a mission is complete?

An automated evaluator, which Replit calls an independent AI judge, checks the evidence against your agreed success criteria. It either accepts completion or returns feedback on what remains. Replit advises you to review the result yourself.

Sources

Image: Replit logo, public domain, via Wikimedia Commons, resized. Replit is a trademark of Replit, Inc. FreeAnswers is not affiliated with or endorsed by Replit.

Related answers

Leave a Comment