defract › blog

Kiro vs Paradev: spec-driven IDE or parallel agent fleet?

2026-11-05 7 min read

Kiro and Paradev both want to be your primary environment for agentic development. They are funded, actively shipping, and both show up when developers search for structured ways to run Claude Code or other AI coding tools at scale. Beyond that, they are almost opposite bets.

Kiro is a structured IDE that generates a spec before it writes code. Paradev is a fleet manager that runs as many agents as you can point at a problem. One constrains how you build; the other maximizes how many processes you run. The choice usually depends on which bottleneck you are actually hitting.

Kiro: spec-driven IDE on Amazon Bedrock

Kiro is built on Code OSS (the open-source VS Code base) and runs on Amazon Bedrock. The core concept is that agents should not start writing code until they have a spec: a structured artifact that captures what to build, why, and how. The IDE generates that spec from your description, then lets you edit it before anything runs. Only after the spec is approved does Kiro proceed to implementation.

The workflow engine is the most direct lane competition with structured lifecycle tools. Kiro ships three named workflow recipes: investigate (read-only research with no writes), feature-pipeline (multi-stage delivery with approval pauses between steps), and publish-pr (branch, commit, and PR creation with review pass). All three run as background cloud sessions. CLI 3.0 (November 3, 2026) made the V3 agent harness the default, bringing cross-workspace global hooks, persistent memory management, and subdirectory AGENTS.md support. The feature-pipeline recipe pauses at planning and review phases for human input before continuing, which is Kiro's version of a gate.

The teams layer is Crew 0.9.0 (October 2026): crewmate profiles with per-role memory and permissions, dynamic dashboards (standup, roadmap, kanban, project reports), and token-usage tracking across the team. Kiro added cross-session persistent memory (IDE 1.4.0) so agents carry project conventions and past decisions into new sessions without re-briefing. "Approve for me" (CLI 2.36.0) lets Kiro auto-approve routine tool actions while flagging risky ones, similar in concept to tunable autonomy presets.

Pricing runs on Bedrock interaction credits, not your own API subscription. Free tier: 50 interactions per month. Pro: $20 per user per month (1,000 interactions). Pro+: $40 per user per month (2,000). Pro Max and Power tiers go to $100 and $200. Overage is charged per interaction.

Where Kiro stands out:

  • Spec generation before code. The IDE authors a structured spec from your description before writing a line. That artifact is editable and version-controlled, so intent is captured and reviewable, not implicit.
  • Structured workflow recipes with approval gates. The feature-pipeline recipe runs multi-stage work with pause-and-approve checkpoints, not a single prompt-to-output run.
  • Crew teams layer. Per-role memory and permissions, dashboards, and token tracking are live in Crew 0.9.0, not on a roadmap. Teams that want structured coordination in a VS Code-compatible environment have something to evaluate today.
  • Familiar editor surface. VS Code settings, extensions from Open VSX, and an interface most developers already know. The adoption barrier is lower than a purpose-built non-IDE tool.

Where Kiro adds friction:

  • Bedrock credit economics. You are not using your own Anthropic subscription. Credits are metered per interaction; heavy usage on Pro+ at 2,000 interactions for $40 is about $0.02 per interaction. At scale this is more expensive than BYO-subscription alternatives, and the billing model is different from what Claude Code users are used to.
  • Vendor-defined workflows only. The three bundled recipes cover investigate, feature delivery, and PR creation. You cannot author your own process or change the stage sequence; the lifecycle is Kiro's, not yours.
  • AWS ecosystem fit. Kiro runs best if you are already in AWS or open to Bedrock models. Teams with existing Claude API or Anthropic subscription agreements get no cost benefit here.

Paradev: 100+ agents, ten coding tools, $11.5M behind it

Paradev rebranded from Superset on October 26, 2026. The branding changed (superset.sh redirects to paradev.com); the product, accounts, and pricing did not. Version 2.0.0 shipped with the rebrand, backed by an $11.5M seed led by Union Square Ventures. At the time, the platform had 145,000 developers across 180 countries and 1.85 million agent workspaces created.

The core proposition is parallelism at volume: run 100+ CLI coding agents simultaneously, each in an isolated git worktree, across any supported agent CLI. Paradev supports ten: Claude Code, Codex, OpenCode, Cursor Agent, GitHub Copilot CLI, Mistral Vibe, Kimi Code, Grok CLI, Gemini CLI, and others. Bring your own API subscription on every plan. For teams that run multiple coding tools or want to compare outputs across models, the routing layer is a genuine capability Conductor and most lifecycle tools do not offer.

Desktop v2.0.0 moved the app home directory from ~/.superset to ~/.paradev and shipped a redesigned shell with rail navigation, title-bar tabs, and pane cards. The October 25 plugins release made plugins available to all accounts, activating across all agents on all machines once installed. For teams standardizing a linter, formatter, or custom tool across Claude Code and Codex sessions simultaneously, the cross-agent plugin layer is a differentiating capability. Native IDE integrations extend the workspace into VS Code, Cursor, JetBrains, and Xcode.

Pricing is BYO-subscription with four tiers: Free ($0, 1 user, local workspaces), Pro ($16 per user per month), Team ($54 per user per month, cloud workspaces in beta), Ultra ($200 per user per month), and Enterprise (custom). Paradev Pro at five users costs $80 per month total. SOC 2 Type II certification (August 2026) is live, relevant for enterprise procurement.

Where Paradev stands out:

  • Scale ceiling. 100+ simultaneous agents is the highest practical ceiling in the category. For test generation sweeps, multi-path prototyping, or documentation across a large codebase, nothing else reaches it.
  • Multi-tool routing. Ten CLI agents from one interface, with per-agent plugin consistency. Teams running Claude Code and Codex or experimenting across models manage both from the same workspace.
  • BYO-subscription economics. Every plan uses your own API keys, so pricing reflects your existing model contracts. No interaction credit overhead on top of what you already pay.
  • Per-seat pricing scales. Five developers on Pro pay $80 total. Paradev Pro is cheaper than most per-seat lifecycle tools at team scale.

Where Paradev adds friction:

  • macOS only (Linux experimental). Windows support is not available. Teams with Windows developers cannot use Paradev uniformly.
  • No process layer. Paradev dispatches agents efficiently but does not define what they build or in what order. There is no spec generation, no approval gate between stages, no design checkpoint. It runs the work you already know you want; it does not help structure what that work should be.
  • Cloud workspaces gated to Team tier. Remote cloud execution (agents running when your machine is off) is $54 per user per month and is in beta. Local-only on Free and Pro.

Which to pick

Pick Kiro when: you want a spec generated before agents write code, your team is in the AWS ecosystem or open to Bedrock billing, you need a teams layer with dashboards and per-role memory today, or you want structured workflow recipes (investigate, feature-pipeline, publish-pr) with approval gates built into an editor you already know. Kiro CLI 3.0 and Crew 0.9.0 together make it a meaningfully different tool than it was six months ago.

Pick Paradev when: your bottleneck is throughput rather than structure, you run multiple coding tools and want them managed together, BYO-subscription economics matter at your usage level, or you need 20+ parallel agents in a single session. The $11.5M seed and daily release cadence since the rebrand mean the platform will ship faster. The multi-tool routing and cross-agent plugin layer are capabilities nothing else in this category matches.

These tools are also not mutually exclusive. Paradev supports Claude Code, Codex, Cursor Agent, and others as routing targets, so a team could use Kiro for structured spec work and Paradev for parallel fleet dispatch from the same codebase. That said, most developers pick a primary environment rather than stacking two. For a broader look at where parallel runners, spec tools, and lifecycle orchestrators compare on task complexity and agent count, the agentic coding patterns that actually scale guide has the full breakdown. The earlier Conductor vs Paradev comparison covers the other major parallel runner in this space.

The process layer neither covers

Kiro adds structure to HOW agents build (spec first, then workflow steps). Paradev adds scale to HOW MANY agents run. Neither answers a different question: WHO decides what gets built, WHEN to stop, and WHETHER the output is correct before it ships. That question is not about spec format or agent count. It is about the gatekeeping layer above dispatch: a design checkpoint that verifies direction before implementation starts, a review gate where agents check each other's work, and a human sign-off at each stage transition.

That is the layer defract occupies: a gated lifecycle (story, design, architecture, implementation, review, release) running on top of Claude Code with a visual design stage and specialist review agents that audit each other before you approve. It is not a spec generator like Kiro or a fleet manager like Paradev. It sits at the stage above both, defining the process that runs before agents start and after they finish. For teams where coordination and correctness are the bottleneck (not throughput or spec generation), why AI coding agents need hard stage boundaries explains the problem it is solving.

defract

a structured lifecycle for your parallel Claude Code agents. free, no caps, no signup.