AWS Kiro adds on-demand "powers" for cleaner, context-aware AI coding
AWS introduced new "powers" for Kiro, its agentic coding environment, and they solve a real problem: tool bloat killing context. Instead of loading every tool up front, Kiro now loads the right capability only when you ask for it. That keeps tokens down and leaves more room for actual code and reasoning. Kiro went GA last month and continues to lean on a spec-first workflow before it writes a line of code.
Why this matters for developers
Every tool you enable burns context before you even start. Load too much, and runs slow down and costs go up. Powers let you activate what you need, when you need it-then deactivate as tasks change. Cleaner context, tighter feedback loops.
How "powers" work
Powers are a unified way to dynamically load tools, Model Context Protocol (MCP) configs and custom instructions based on your prompt. Mention "database," and Kiro can detect a Neon setup and bring it online. Say "payment" or "checkout," and the Stripe power spins up to generate code against their APIs. When the task shifts, that power turns off, freeing memory for what's next.
MCP without the context tax
MCP servers standardize how agents reach external tools and data-great for real work like tests, debugging and service scaffolding. But traditionally they load everything up front. A Figma server might expose eight tools that eat ~12,000 tokens; a Postman server can add 120+ tools. Powers flip that model by pulling only what the current task needs.
If you're betting on agents for real dev work, this is the difference between "demo nice" and "production reasonable."
Getting and managing powers
Developers can browse and install powers inside Kiro or from kiro.dev. At launch, partners include Figma, Stripe, Neon, Supabase, Netlify and Datadog. You can also import community powers from GitHub or build private ones for internal workflows.
Where you can use them
Today: powers run inside the Kiro environment. Next up: support across the Kiro CLI and popular AI coding tools like Cline, Cursor and Claude Code. AWS says the goal is compatibility with any AI dev tool.
Practical tips to roll this out
- Start lean: install a few high-signal powers (e.g., repo tooling, tests, your primary DB, your payment stack).
- Map intents to powers: define trigger phrases and conventions so teams know how to "call" a capability.
- Scope credentials: least privilege per power; isolate environments (dev/stage/prod) with clear guardrails.
- Measure token spend: track runs before/after powers to quantify savings and reasoning gains.
- Version and pin: lock power versions for reproducibility; promote updates through CI like any dependency.
- Document failure modes: define fallback commands when a power isn't available or misfires.
- Security review: audit any power that touches secrets, infra or production data.
Standardizing agent access
Powers lean on the Model Context Protocol to avoid one-off integrations. That gives teams a cleaner way to expose repeatable tasks-run tests, seed data, generate service scaffolding-without stuffing the context window with everything at once. The payoff: fewer tokens burned on static definitions and more on reasoning and code.
The takeaway
Kiro's powers push agentic coding toward practical, cost-aware workflows. Load tools just in time, keep memory clear, and let the model focus on the job. If you've hit the ceiling on context size or watched token bills spike, this is worth piloting on a real service.
Want to skill up your team for AI-assisted coding?
Explore hands-on paths here: AI Certification for Coding.
Your membership also unlocks:
