Replit Fresh off a high-profile coding fiasco, Replit is making a big bet on a new guardrail: an AI coding agent that scrutinizes and repairs its own work. It’s called Agent 3, and it promises greased-lightning speed, lower costs and less-risky “vibe coding” for people asking for software due to be built on nothing more than plain English prompts — no need to be an engineer.
What Agent 3 does
Agent 3 translates natural-language instructions into functional projects by architecting, compiling code, executing it, and revising work iteratively until tests pass. Autocomplete does more than autocomplete: It constructs end-to-end features, wires up services and can even spin up helper agents that plug into Slack, Telegram or Gmail.
Replit claims the new agent is three times faster and 10 times more cost-effective than “Computer Use” models that mimic a person controlling a desktop. The agent focuses on mainstream languages and frameworks, and it’s designed to make sure that users have previews, diffs and explainers at every step of the way.
How the self-checking loop functions
Agent 3 is based on a (plan–build–test–fix) development cycle. It runs the code you write, provisions the program in the sandbox, captures runtime errors, and fixes the failing bits. Repeatedly it tries this loop until the tests and the health check are green. Think of it as a junior developer whose push you won’t accept until it passes cleanly.
To increase confidence, the agent relies on multiple signals – unit and integration tests where present, linter and type-checker output, basic static analysis to identify dangerous calls or missing dependencies. It also looks at diffs before making changes, so if something looks off, a reversion is easy.
Replit’s leaders said the company would harden isolation and add safer defaults to development environments as a result of the previously reported incident, which they said had been widely discussed. In practice, that means tighter file permissions, clearer prompts before destructive actions, and a focus on writing tests alongside features, so that an agent has something conrete to check for.
Why this matters now
Agentic coding has been one of the relatively few use cases of generative AI that is consistently able to result in productivity improvements. In GitHub’s studies with developers, AI-augmented individuals finished tasks around 55% faster, and internal enterprise rollouts frequently see significant gains on boilerplate or repetitive tasks.
The market is moving quickly. A few months ago, Replit rolled out a new $250 million raise that pegged its value at around $3 billion, while reports from Reuters claim that Cognition and Cursor are in the neighborhood of $10.2 billion and about $10 billion. Separately, Anthropic and others have published research cautioning that autonomous agents can deceive and take risks in uncertain objectives—highlighting that self-checking and sandboxing are not “nice-to-have[s].
How to test Agent 3
Begin by logging into Replit and opening a new project. Choose a template (i.e., a Python web app), enable Agent 3 in your workspace, and describe your request in a single sentence or brief. The agent will suggest a plan, create files, execute the app and display a preview.
If the run is not successful, Agent 3 examines logs, will enact a fix, and ask to re-run the checks until basic checks pass. You can accept or reject changes, request tests, or ask it to refactor for readability. If your project depends on external services, drop in API keys via environment settings and get the agent to wire integrations like Slack or Gmail.
Pricing is simple. There’s a free Starter tier trial for Agent 3, while access to all features via the Core subscription is around $20 per month.
Teams and Enterprise plans offer policy controls and collaboration features for organizations that seek guardrails and auditability.
For safety, fork key repos before you play, run the agent in a branch by itself, and require tests for any merge. Keep human-in-the-loop reviews for production paths—particularly database migrations, permission changes, and infrastructure adjusts.
Early takeaways
Agent 3’s self-checking loop won’t stamp out bugs, but it does significantly reduce the time that passes between idea and working prototype. The agent excels at scaffolding projects, wiring up APIs, and grinding through boilerplate edits, letting people make final architecture and safety calls.
If Replit’s quarantine and review functions withstand real-world usage, the company’s bet could redeem itself after its early stumble—and help nudge agentic coding from novelty to dependable utility for both startup and enterprise use.