Scaling the Claw: Skills, Shell, CodeAct and Background Agents

dotnet explains four additions that make an AI agent more practically useful: skills for loading domain knowledge on demand (including centrally managed skills in Azure AI Foundry), constrained shell access for file work, a CodeAct sandbox for writing and running code, and background agents for parallel research.

Overview

The video describes four capabilities that improve an agent “harness” so it can handle real tasks more reliably:

Skills (domain knowledge on demand)

• Skills package domain knowledge (examples mentioned: valuation and risk scoring) into files.
• The agent discovers and loads these files only when a request calls for them.
• Skills can be managed centrally in Azure AI Foundry so they can update without requiring an application redeploy.

Shell access (constrained file operations)

• Shell access allows the agent to reorganize a messy folder of trade confirmations.
• The shell is confined to a single working directory to limit what the agent can access.

CodeAct (code-writing and execution sandbox)

• CodeAct provides a sandbox where the agent can write and run code.
• The goal is to let the agent compute answers (for example, arithmetic) instead of guessing.

Background agents (parallel work)

• Background agents allow the system to research several tickers at once.
• Results can then be pulled back together into a combined output.

Links

• Series info: https://aka.ms/FMTA/series
• Resource: https://aka.ms/AgentHarness