Proposal: Upstream "Intent Contracts", "Execution Profiles", and "Autonomous Research" as Kiro Powers/Primitives

[chernistry]

Problem Statement
Kiro has excellent foundational primitives: Steering (.kiro/steering/) sets global project context, and Hooks provide powerful runtime control.

However, based on my collaborative agent experiments ("Kotef"), I've found that global steering isn't enough to prevent "agent drift" during specific tasks. Agents often:

1. Suffering from Task Creep: Steering says "Prefer small changes", but the agent decides this specific ticket requires a total refactor.
2. Getting stuck in Verification Loops: Spending hours debugging a minor CSS fix because they are treating it with the same rigor as a core API change.
3. Confident Hallucination: Using "remembered" patterns instead of looking up the latest documentation for a specific library version.

Proposed Solution: Patterns on top of Kiro Primitives
I propose standardizing three "Agent Design Patterns" that leverage Kiro's existing architecture to solve these problems:

1. Intent Contract (Dynamic Steering)

   • Concept: Instead of just static .kiro/steering/product.md, the agent generates a temporary, task-specific Intent Contract (JSON) at the start of a session.
   • Implementation: A custom agent that defines Non-Goals (e.g., "Do not touch src/legacy") and Appetite (e.g., "Small - max 5 tool calls") for just this run.
   • Kiro Fit: This acts as a "Session Steering" layer.

2. Execution Profiles (Hook Configurations)

   • Concept: A standard library of Hooks (Stop, PreToolUse) that toggle verification rigor based on the task.
   • Implementation:
     • Strict Profile: Enforces git checkpoints on green tests (Fail-Closed).
     • Fast Profile: Runs standard linting but allows minor failures.
     • Yolo Profile: Minimal checks for prototyping.
   • Kiro Fit: Pre-packaged post_tool and stop hook configurations that users can swap via CLI settings.

3. Autonomous Research (Agentic Loop)

   • Concept: A "Look before you leap" workflow that uses Kiro's search and fetch tools before attempting code.
   • Implementation: An agent loop that scores research quality (Relevance, Recency) and refines queries until it has a solid plan, preventing hallucination.
   • Kiro Fit: A specialized "Research Agent" or thinking model prompt strategy.

Evidence / Links
I have implemented these patterns in a standalone LangGraph agent ("Kotef") and verified their impact.

• Intent Contract Logic: intent_contract.ts
• Execution Profiles: profiles.ts
• Research Node: researcher.ts

What I'm willing to contribute

• Porting the "Intent Contract" logic into a Kiro Custom Agent template.
• Sharing the "Deep Research" prompt patterns for Kiro's thinking models.
• Documenting "Execution Profiles" as a best practice for Kiro Hooks.