SKAP

Skill Adapter Protocol

July 2025

Skill Adapter is a lightweight protocol for adding State Management to AI agents. While Skills teach an agent how to perform a task (Capability) and Rules tell an agent what not to do (Constraints), Skill Adapters tell an agent where they are in a workflow (Orchestration).

Why Use a Skill Adapter?

Eliminate Loops

Prevents agents from getting stuck optimizing a single step forever (e.g., trying to bypass a paywall for 20 minutes)

Enforce Persistence

Ensures the agent completes long-horizon tasks (50+ steps) without "forgetting" the goal

Dynamic Logic

Allows the agent to change behavior based on context (e.g., "Be fast during discovery" vs. "Be thorough during verification")

SKAP 1.0 vs 2.0

Previous Version: SKAP 1.0 (UI Mapping)

The original SKAP focused on UI mapping—teaching agents where UI elements are located and how to interact with them. This approach achieved 33% better performance on MiniWoB++ benchmarks, validated across 2,000+ episodes.

0.64 average reward
12% better than Gemini-2.5-Pro
p < 0.05 statistical significance

Current Version: SKAP 2.0 (State Management)

SKAP 2.0 shifts focus to State Management. Instead of just mapping UI elements, it tells agents when and why to use specific skills based on workflow state.

The Problem: Agents follow instructions perfectly but lose track of why they're doing something. They optimize steps without understanding the goal.

The Solution: State-based execution. The agent knows "I am in [DISCOVERY] state, so speed matters more than accuracy" vs "I am in [VERIFICATION] state, so accuracy is critical."

Note:SKAP 2.0 has not yet been benchmarked. The 33% improvement metrics refer to SKAP 1.0 (UI Mapping approach).

Why State Matters

The shift from UI Mapping to State Management solves what I call the Reasoning Tax—the computational cost of making agents "figure out" what to do at every step.

Instructions
Mechanical steps that are binary and static.
"Click the Deploy button"
"Write Python code"
Intent
Desired outcomes that are fluid and contextual.
"Deploy safely"
"Write clean code"

The gap between Instructions and Intent is State. Agents fail when they follow instructions perfectly but lose the context needed to understand intent. State Management bridges this gap by making workflow position explicit.