What Replit shipped

On 17 November, Replit announced Missions. A mission hands one goal to a group of cooperating agents. Those agents keep working in Replit's cloud until they meet an agreed bar, run out of budget or need a human decision. Replit calls the product an agent swarm harness. A harness is the orchestration code around a model: it decides what each agent does, which tools each agent can call and when work stops. According to the launch post and the Missions product page, builders can start a mission from the "New" menu in the sidebar, with the /mission command, or from a suggestion the agent makes.

Most coding agents run one session against one prompt. Missions puts several sessions under a written agreement that sets the spending limit, the permissions and the success criteria before any work starts. The agreement is the part builders should study closely, because it controls both the cost and the moment the mission counts as done.

How a mission runs

Setup produces what Replit calls a mission agreement. The launch post says the agreement records the spending cap, the actions that need human permission and the tradeoffs the agents may make. Replit's example of a tradeoff is offering paid pilots. During setup, Missions also asks questions to pin down success. For Replit's sample goal of winning the first 10 paying enterprise customers, the agent asks what counts as an enterprise customer and whether paid pilots count.

After sign-off, agents work in parallel through the connectors the user grants. The product page lists GitHub, Linear, Notion, Google Drive, Figma, HubSpot and Airtable. Software work comes back as pull requests and reviewable files. Spend is tracked per step, with steps such as research, build, test and review. Messages the user sends during a mission count toward the mission's budget.

The stopping rule is what separates Missions from an ordinary long-running agent loop. A separate AI judge, which Replit also calls an eval judge, checks the output against the agreed criteria. When something is missing, the judge sends specific feedback back to the agents. According to the product page FAQ, a mission continues until the criteria are met, the budget runs out or the mission needs the user's input. The launch post says that if the budget runs out first, the outcome is left unresolved. The product page says any budget increase needs the user's approval. Users can get progress reports immediately, as a daily digest or weekly.

What Replit's own numbers show

Replit backed the launch with internal tests. On 13 NumPy optimisation tasks, the launch post reports speedups of up to 4x for one agent on Auto (Max), 17x for 16 agents and 19x for 32 agents. On Auto (Power), the same three set-ups reached 2x, 7x and 9x. Going from 16 to 32 agents added much less than going from one agent to 16. On this task set, the gains shrink well before a swarm gets large.

The bigger demonstration is DataCat, an observability platform. Replit says Missions built DataCat in about two days within a $3,000 cap. The DataCat repository is public under Apache 2.0. DataCat stores data in PostgreSQL or ClickHouse and accepts OpenTelemetry and Datadog-compatible agent traffic. Replit compared DataCat with a build from the same prompt using Claude Code's /goal command. On Postgres, DataCat ingested 49,000 traces per second against 9,800 for the Claude Code build. DataCat also passed 51 of 56 API compatibility checks, against 45. Replit's post also says the Claude Code build covered more product areas and ingested logs faster.

Replit lists its own caveats. The comparison used one run per set-up, scored only the product areas both builds shared, and doesn't show how DataCat performs against the hosted service it imitates. Cocoloop's write-up points out that Replit didn't disclose the Claude Code build's budget, time or prompts. We'd treat the figures as a vendor demonstration.

Costs and unknowns

Neither Replit page publishes a price for Missions. Replit's AI billing docs describe effort-based pricing for Agent work. Each request ends at a checkpoint, and the charge reflects the effort that went into the request. Power and Max modes are billed per request, and third-party models are charged at the provider's public API rate and deducted from credits. The billing docs don't mention Missions. We don't know how a swarm's parallel sessions are metered, how many agents each plan can run at once, or whether the plan limits on background tasks apply to Missions. The docs set those limits at zero on Starter, one on Core and 10 on Pro.

Account-level controls sit outside the mission agreement. The managing spend docs say organisation budgets must be set in increments of $500. Hitting the service shutdown limit suspends usage-based services until the budget goes up or the next billing cycle starts. Replit's pages quote dollar figures without naming a currency, so NZ teams should check how their own account is billed. We'd set an account usage limit as well as a per-mission cap, so a mistake in a mission agreement can't run past the account limit.

The judge is the other open question. An AI judge can only enforce the criteria written at setup. Some goals Replit showcases, such as winning customers, are hard to check automatically. Replit hasn't published how often the judge accepts work that a person would reject. The product page says a human reviews the finished work and its evidence before release where the agreed boundaries require that review. We'd keep the human review on for anything customer-facing. The pages also don't say where the cloud agents run. Teams that would connect a Google Drive or HubSpot account holding customer data should find out before granting access.

What a builder could do with Missions

Replit's strongest examples have measurable success conditions: a benchmark speedup, a compatibility suite, a latency target. The templates on the product page include API latency, a support assistant and a Rust rewrite. A sensible first mission for an Auckland team is a well-tested internal service with a clear performance target and a small cap. Running the same task in the team's usual coding agent gives a direct comparison of spend and output.

Success criteria work best when written as checks a script could run, such as a test command that must pass or a p95 latency under a stated number of milliseconds. The judge then has something concrete to enforce. We'd limit connectors to the repository and the issue tracker for early missions, and keep outbound email and CRM writes behind approval. If members run a mission against a real codebase, we'd like to see the cap, the actual spend and the result shared on the ASAI projects page.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member