Kiro: A new agentic IDE

Original Article ↗ Hacker News Discussion ↗

What Kiro Is and How It’s Built

  • Agentic IDE built as a VS Code fork, powered behind the scenes by AWS Bedrock (Claude Sonnet 3.7/4).
  • Offers chat, spec mode, and “agent hooks” that can run multi-step workflows (e.g., updating tickets, syncing with external tools).
  • AWS product but deliberately branded and hosted somewhat separately; uses AWS legal terms and IAM Identity Center for enterprise login.

Spec‑Driven Development & Steering

  • Core differentiator is “spec-driven development”: three main files – requirements, design, tasks – plus “steering” rules in .kiro/steering.
  • Requirements enumerate edge cases; design contrasts current code vs requirements; tasks break work into LLM‑sized chunks and track progress.
  • Users report it adds structure to “vibe coding” and scales better on medium–large codebases, though some find it verbose and over-complicating solutions.
  • Specs are currently mostly static; some use them as an append-only design history rather than a single canonical doc.

Comparisons: Cursor, Claude Code, CLI Tools

  • Many see it as “another VS Code AI fork” in an already crowded field (Cursor, Windsurf, Zed, etc.).
  • Debate over IDE vs CLI/TUI: IDEs provide richer context (LSP, problems panel, open files) and lower tool latency; CLIs are editor-agnostic, scriptable, and easy to run in CI.
  • Some argue similar workflows can be achieved today via Claude Code plus rule files (CLAUDE.md / AGENT.md) or tools like Cline/Roo/Aider.

Pricing, Interactions, and Data

  • Priced by “agentic interactions” (human-initiated runs) rather than tokens; Pro/Pro+ include 1,000–3,000 interactions with overage at $0.04 each.
  • Discussion on whether these limits are generous or constraining, compared to Claude subscriptions and Amazon Q Developer pricing.
  • Free/preview tier may use content to improve models unless opted out; paid tiers and Q Developer-linked usage are excluded from FM training. Some distrust whether such promises are verifiable.

Performance, Resource Use, and Bugs

  • Initial indexing and plugin import can cause high CPU/RAM; large projects may trigger ongoing re‑indexing.
  • Multiple reports of login/SSO failures (Google/GitHub), extension host crashes on Linux, terminal windows popping open unexpectedly, and high CPU on large repos.
  • Lacks devcontainer support due to dependence on proprietary VS Code remote extensions; some key Microsoft extensions are incompatible by license.
  • Several users find it slower and more brittle than Claude Code/Cline on real tasks, and rate-limited under heavy use.

Adoption, Lock‑In, and Workflow Concerns

  • Strong reluctance to switch IDEs repeatedly; many prefer editor-agnostic or plugin-based approaches (JetBrains, Emacs, Neovim, Helix, Aider).
  • Frustration with fragmented “rule files” across tools; some call for a standard like AGENT(S).md, others argue it’s too early to standardize.
  • Skepticism rooted in prior Amazon Q experiences (seen as half‑baked) and perception that VS Code forks mainly serve as data funnels and lock‑in plays.

Broader Reflections on AI Coding

  • Several comments argue the real value is in rigorous specs and architecture, with LLMs handling “easy” implementation work.
  • Others worry that agentic flows push developers into PM‑like roles, risk environment corruption without proper sandboxing, and still require significant oversight.
  • Mixed reports on effectiveness: some teams claim clear productivity gains; others find Kiro (and similar agents) fragile on nontrivial tasks or complex environments.