The eleven tools we've covered are the established names. Audiences found, positioning refined, in some cases gone mainstream. But this category moves faster than any release cycle. The five platforms below are newer, bolder, more experimental — and worth understanding before they become the story everyone's talking about.

The through line: the industry is shifting from AI-as-assistant to AI-as-agent. The earlier tools help you code faster. These tools are starting to ask whether you need to be involved in the coding step at all.

---

Kiro

Amazon's spec-first IDE that brings engineering discipline to vibe coding — without killing the vibe.

---

What This Is

Kiro is AWS's AI-powered IDE with a different philosophy from every other tool in this guide. Most AI IDEs start with "describe what you want, we'll write code." Kiro starts one step earlier: "let's write a spec first, then build from it."

Built on VS Code (like Cursor), so the editing experience is familiar. The differentiator is spec-driven development: before any code, you work with Kiro to produce a specification — user stories, data models, API contracts, task breakdowns. The AI uses that document as its north star throughout the build. The spec keeps the AI grounded. Instead of drifting across sessions or forgetting what the feature was supposed to do, Kiro anchors every action to a document you both agreed on upfront.

Tightly integrated with AWS — Lambda, DynamoDB, S3, Cognito, API Gateway. That's both its biggest strength and clearest limitation.

Who It's For

AWS-native teams. If your stack is on AWS, Kiro feels almost supernaturally capable.

Teams that want process, not just speed. The spec-first model directly addresses vibe coding's biggest team problem: code nobody else understands, with the spec living only in someone's head. Engineering managers worried about quality and handoff find Kiro reassuring.

Experienced devs bridging into AI. Kiro assumes you can read a spec and push back when something's architecturally wrong.

What Actually Matters

• Spec files as the engine. Every code change traces back to the spec. Update spec, AI adapts.
• Steering files for persistent project context — including AWS account details, service config, team standards.
• Hooks that auto-trigger on file changes — save an API route, tests update; modify a schema, docs update.
• AWS CDK / CloudFormation scaffolding that's production-grade, not prototype-tier.

Cost

Out of preview and generally available, with real pricing. Credit tiers: Free (50 credits/month), Pro $20, Pro+ $40, Pro Max $100, Power $200; extra credits $0.04 apiece. Fine print: spec-mode requests burn more credits than vibe-mode ones — the structure costs extra. "Kiro on the Web" is in preview for paid tiers.

Best At

• Web apps: Five stars if you're on AWS.
• Mobile: Three stars. Functional, doesn't leverage Kiro's edge.
• Local/scripts: Three stars. Spec-first overhead isn't worth it for small tasks.

Where It Bites

The spec-first workflow adds upfront time. Pure exploration mode? Friction. Non-AWS users get a capable tool without specialization.

---

Google Antigravity

Google's agent-first development platform — and the official heir to Firebase Studio, which just joined the Google graveyard.

---

What This Is

First, a eulogy. Earlier editions covered Firebase Studio here, and I warned then that Google has a history of deprecating developer tools. Called it. Firebase Studio signups closed June 2026; full sunset March 22, 2027 — deployed apps keep running, the IDE goes dark. The official migration path is Antigravity, conveniently also Google's actual bet.

Antigravity launched November 2025 alongside Gemini 3 as an agent-first VS Code fork. At I/O 2026, version 2.0 pivoted from "editor with AI" to full agentic platform: desktop app, CLI (it replaced the Gemini CLI outright), and an SDK for orchestrating teams of agents in parallel. You don't sit in a chat pane — you dispatch agents, and they come back with evidence.

Who It's For

Gemini ecosystem users — including everyone Firebase Studio just orphaned. This is where Google wants you.

Developers who want autonomous agents with receipts. Antigravity's signature move is verification artifacts, not just diffs.

Anyone who wants a frontier agentic IDE for free. The individual tier costs nothing and the quota is genuinely usable.

What Actually Matters

• Agent Manager. Multiple agents on multiple tasks in parallel, mission-control style. No babysitting one chat.
• Artifacts as evidence. Plans, screenshots, browser recordings of the thing actually working — verification you can review, not vibes to trust.
• Browser control built in. Agents test their own work in a real browser before claiming success.
• Model-flexible. Gemini 3 is the headliner, but non-Google models are supported too. Rare humility, for Google.

Cost

Free for individuals with a metered quota. Higher limits ride on Google AI subscriptions — Pro around $20/month, Ultra tiers from roughly $100 to $200/month with 5x–20x the quota.

Best At

• Web apps: Four stars. Agents that test in a real browser are a genuine edge.
• Mobile: Three stars. Flutter is decent; not the home territory Firebase Studio had.
• Local/scripts: Four stars. The CLI made this a first-class citizen.

Where It Bites

It's young; workflows you learn this quarter get rearranged next quarter. Free-tier rate limits pinch at peak hours. Unsupervised parallel agents produce parallel messes. And the elephant: you're now two Google deprecations deep in this product line. Fool me twice.

---

Devin

The first AI that applies for the job, doesn't ask for standups, and ships code while you sleep.

---

What This Is

Devin is not an AI coding assistant. It is an AI software engineer.

Every other tool in this guide helps you write code faster. Devin takes the spec, runs the entire engineering loop itself: plans, writes, tests, reads errors, fixes bugs, deploys. You assign it a task the way you'd assign one to a junior engineer. It works independently, reports back when done — or when stuck.

Cognition AI launched Devin in early 2024 to a press cycle that ranged from astonished to skeptical. Since then Cognition swallowed Windsurf, rebranded its IDE as Devin Desktop, and raised at a $26 billion valuation on roughly half a billion of ARR. The skeptics got quieter. The honest answer today: reliable on well-specified tasks, less reliable on work requiring judgment or domain knowledge.

That's still meaningful. A task that takes a competent engineer three hours and $500 in salary can be dispatched to Devin for a fraction of the cost, completed asynchronously while you focus on higher-order decisions. The question is always whether you're giving it the right kind of task.

Who It's For

Engineering teams with clearly specified backlog items. Devin's quality scales with task spec quality. Well-scoped tickets, clear acceptance criteria — the kind a good team writes anyway.

Technical founders with a feature backlog. Execute against it overnight. Review the PR in the morning. This is asynchronous leverage that changes what one person can ship.

NOT for non-technical builders. You need to evaluate output, catch what it got wrong, give precise feedback. Without that judgment, you're accepting code you can't verify.

What Actually Matters

• Autonomous execution. Browser, editor, terminal in a sandbox. Iterates on its own — minutes to hours of work per task.
• GitHub integration. Pull from issues, work on a branch, open a PR. Assign Devin like a team member.
• Async reporting. Status updates while it works. Notification when done. Tells you where it got stuck if it does.
• Project memory. Correct an approach once — "we use hooks, not class components" — it remembers for next time.

Cost

No longer the luxury item it launched as. Mid-2026 brought normal tiers: free, Pro $20/month, Max $200/month, Teams $80/month plus $40 per seat — usage credits on top for heavy agent runs. Entry is trivial; spend scales with how much autonomous work you dispatch, and heavy users still land in the hundreds per month. Expensive until compared to a contractor at the same hours. Check cognition.ai for current rates.

Best At

• Web apps: Four stars. Backend APIs, data models, integration, debugging — solid. Watch for architectural drift.
• Mobile: Three stars. Better at backend logic than UI work.
• Local/scripts: Four stars. Probably its most reliable use case.

Where It Bites

Requires technical oversight not every user can provide. Devin is confident about its output — won't always flag when something's wrong. Cost structure punishes poorly specified work — you burn compute dollars without useful output. Treat it like a capable but expensive contractor who delivers exactly what you asked for, even when what you asked for wasn't quite right.

---

Manus

A general-purpose AI agent that browses, codes, reasons, and ships — without being told which tool to use next.

---

What This Is

Manus is the outlier in this chapter. Every other platform here is, at its core, a coding tool. Manus is an AI agent that happens to be capable of coding — among many other things. It can research a topic on the web, synthesize findings into a document, write and execute Python, manage files, fill out forms, interact with external services — all within a single coherent task.

The positioning from its creators: "the first general AI agent." You tell it what outcome you want. It figures out the path.

For builders, the practical application is significant. Manus can take "research the top five competitors in project management, build a comparison table as a web page, deploy it somewhere I can share" — and actually complete all of it in sequence, including the parts that cross research, writing, coding, and deployment. No other tool in this guide does all those steps in a single autonomous run.

Who It's For

Founders and product people who need artifacts, not just code. Reports, working prototypes, deployed demos where the path crosses domains.

Non-technical users who want to delegate entire tasks. Lowest-friction entry point in this chapter for someone without a technical background.

Developers exploring what autonomous agents can actually do. Manus is a useful benchmark for AI agency in practice — what works, what gets stuck, where the technology is heading.

What Actually Matters

• Open-ended task planning. Produces a plan before executing. Review it, modify it, approve it. Plans can be ten steps deep across web research, files, code, external services.
• Multi-modal action. Browse, read/write files, execute code in a sandbox, hit APIs, interact with web UIs. Crosses domains in a single run.
• Persistent context. Mid-multi-hour task and you close the window? It picks up where it left off.
• Lightweight web deployment. Deploys simple pages and tools to public URLs as part of a task. Closes the loop from idea to shareable link in one run.

Cost

Free tier with a daily credit refresh — roughly one real task a day. Paid plans run $20/month (4,000 credits) to $200/month (40,000), credits-based: purchase, consume per task complexity. Check manus.im for current rates.

Best At

• Web apps: Three stars. Great for simple tools and informational pages. Real apps with auth and state — go elsewhere.
• Mobile: Two stars. Wrong tool.
• Local/scripts: Four stars. "Pull my Analytics data, format as a weekly report" — squarely in its zone.

Where It Bites

The generalism is also the limitation. Not the deepest coder, not the best researcher, not the most capable deployer — a capable generalist. Specialists outperform in their domains. The web browsing means it sometimes cites outdated pages. Credits-based pricing produces sticker shock on long, complex tasks if you're not watching usage.

And a corporate soap opera in the background: Meta announced a roughly $2 billion acquisition in late 2025; Chinese regulators ordered it unwound in April 2026; still unresolved. The product keeps shipping, but keep an exit plan handy.

---

OpenAI Codex

OpenAI's coding agent rides shotgun with your ChatGPT subscription — and ships pull requests while you're in a meeting.

---

What This Is

Codex is OpenAI's agentic coding play, and distribution is the story: it comes bundled with the ChatGPT subscription hundreds of millions of people already pay for. Plus subscriber? You already own a coding agent.

Give it a task in plain English and it reads your repository, edits across files, runs tests in a cloud sandbox, and opens a pull request. It lives where you do — CLI, IDE extensions, cloud tasks fired off from your phone. Powered by the GPT-5-Codex model line, iterated relentlessly.

Who It's For

Anyone already paying for ChatGPT. The marginal cost of trying a real coding agent just hit zero.

Teams that want parallel delegation. Kick off five scoped tasks in the cloud, review five PRs after lunch.

Developers in the OpenAI ecosystem. Keeping tooling in the family simplifies life.

What Actually Matters

• Bundled distribution. No new vendor, no new invoice. That's how a tool gets into companies that never approve new line items.
• Parallel cloud tasks. Each gets its own sandbox; none block your machine.
• CLI + IDE + GitHub code review. One agent, multiple surfaces, including automated review comments on PRs.

Cost

Included with ChatGPT plans — Plus $20/month, Pro $200, Business $25 per user — metered in token credits since April 2026. Light use fits inside Plus; heavy daily agent use lands at $100–200 per developer per month. The floor is the subscription you already had.

Best At

• Web apps: Four stars. Multi-file changes, tests, PRs — the core loop is solid.
• Mobile: Three stars. Competent, not differentiated.
• Local/scripts: Four stars. The CLI shines on scoped, well-specified tasks.

Where It Bites

The cloud sandbox stumbles on private dependencies and exotic build setups — expect environment yak-shaving on real codebases. Credit burn is opaque; your first heavy month is a surprise. And it happily generates more code than you have attention to review. The bottleneck moves from writing to reading.

---

Gavin's Rule: Match The Tool To The Task's Shape

The five platforms here do things the earlier tools don't — autonomous execution, spec-driven structure, deep ecosystem integration, generalist cross-domain tasks. Each has a specific shape it fits best.

Before committing, describe your project in one sentence and ask three questions:

1. Does this need a specialist or a generalist?

2. Does it live in one ecosystem (AWS, Google, OpenAI) or cross many?

3. Does it require human oversight at each step, or can it run while I sleep?

The answer to those three questions usually points directly to one of these five — or tells you to go back to an earlier chapter.

---

Bottom Line

The vibe coding market hit $4.7B in 2026. These five platforms represent where the next billion is being built.

• Kiro — for AWS-native teams who want structure to match speed.
• Google Antigravity — for anyone in the Google/Gemini ecosystem, and everyone Firebase Studio left behind.
• Devin — for technical teams with well-specified backlogs who want leverage without headcount.
• Manus — for founders and operators who need outcomes that cross research, writing, and code.
• OpenAI Codex — for anyone already paying for ChatGPT who wants an agent without a new invoice.

None of them replaces the tools in earlier chapters. They extend the frontier.

The question isn't which one is best. It's which one fits the kind of work you actually need to do.

— Gavin