Replit launches Missions, handing agent swarms full goals under spending caps
Replit unveiled Missions on November 17, 2026, letting teams delegate entire outcomes to coordinated agent swarms.

· Last verified: 18 Nov 2026 (Replit Blog)

Summary
- Replit unveiled Missions on November 17, 2026, letting teams delegate entire outcomes to coordinated agent swarms.
- Users set a spending ceiling and specify which actions require their approval before agents act.
- An independent AI judge scores results against success criteria; budget exhaustion halts work unresolved.
The latest
Teams can now hand an entire business goal to a swarm of AI agents rather than directing one agent task by task. Replit announced Missions on November 17, 2026, a tool that takes objectives like winning enterprise customers or replacing legacy software, then plans, builds, tests and iterates toward them. A spending cap and an approval list govern what the agents may do unsupervised.
Details
- The setup: Missions opens by asking clarifying questions to pin down what success actually means, according to Replit's blog post. The company's example: defining what counts as an enterprise customer, and whether paid trials are included. Those answers become the criteria the work is later measured against.
- Budget and guardrails: The agreement fixes a spending ceiling and lists actions needing user sign-off. In Replit's illustration, a $1,000 cap applies, with approval required before contacting prospects, sending offers, altering customer-facing sales material, or making product commitments on the company's behalf.
- How the swarm works: Agents run in parallel rather than sequentially — one researching potential buyers while another builds a demo, per Replit. Results from each stream feed back into the next round, adjusting what the swarm does rather than following a fixed plan.
- Reporting cadence: Users choose what they hear about — blockers, progress, or status — and how often, from immediate alerts to daily or weekly digests. Agents operate in the cloud and Missions reaches out only when it needs input, Replit said.
- Verification and oversight: A separate AI judge checks outputs against the agreed criteria, and agents revise using its feedback. A mission closes when criteria are met; if the budget runs out first, work stops unresolved. Users can review completed work, active tasks and pending approvals. Replit did not publish pricing.
- The DataCat case: Replit used Missions to build DataCat, a self-hostable application monitoring platform, in two days on a $3,000 budget. It caught a deliberately injected database fault in 38 seconds against a 60-second target and passed 51 checks — 36 on data integrity, 15 on alert reliability.
- Load testing: In a three-hour load test, DataCat processed 26 million activity records and 4.4 million event records with no data loss, per Replit. In Node.js testing it handled 39% more requests at 51% lower processor load.
- Head-to-head claim: Built from the same initial prompt, Replit reported DataCat reached 49,000 traces per second against 9,800 for a Claude Code build with Postgres, and 405,100 measurement points per second against 224,600. Replit cautioned these are controlled tests that do not establish performance against hosted or production systems.
- Parallel speedups: Across 13 NumPy tasks, Replit reported Auto (Max) delivering 4x acceleration with one agent, 17x with 16 agents and 19x with 32. On Auto (Power), the same tasks returned 2x, 7x and 9x respectively.
Between the lines
The spending cap doubles as the failure mode: Replit states a mission simply halts when money runs out, leaving the outcome undecided, which makes budget a risk ceiling rather than a performance guarantee. The approval list — covering outreach, offers and product commitments — concentrates human control at the points where an agent's action would reach an outside party.
What's next
Missions is reachable from the New button in the sidebar; enterprises are directed to Replit's sales team. DataCat is open source on GitHub, leaving its benchmark claims open to independent replication by outside developers.
Source: Replit Blog