From Model to Agent: The Agent Framework Harness, Live in C#

4:00 午後 - 5:00 午後 (UTC)

One call to AsHarnessAgent gives us the whole agentic loop. From there we add the first three abilities: a custom get_stock_price tool, the built-in web search for market news, and planning with a todo list plus plan and execute modes. That is enough to turn a vague request like “review my watchlist and suggest something to add” into a tracked, step-by-step plan. We wire the agent to a Microsoft Foundry chat client, run it in the shared harness console, and set up the companion repo so everyone starts on the same SDK version.

4:00 午後 - 5:00 午後 (UTC)

So far the agent can talk about markets but cannot touch anything of yours, and nothing stops it doing something it should not. This episode fixes both. We give it file access so it reads a real portfolio.csv and writes its reports back to disk, then put a place_trade tool behind an explicit human approval. Approving every single read gets old fast, so we add auto-approval rules for the safe calls and one custom rule that clears small trades on its own. We finish with durable memory in two forms, the files the agent curates itself and the facts Foundry extracts automatically, and watch both survive a restart.

4:00 午後 - 5:00 午後 (UTC)

Four additions make the agent genuinely useful. Skills package domain knowledge such as valuation and risk scoring into files it discovers and loads only when a request calls for them, including skills managed centrally in Foundry that update with no redeploy. Shell access lets it reorganise a messy folder of trade confirmations, confined to one working directory. CodeAct gives it a sandbox where it writes and runs code to work out an answer rather than guessing at the arithmetic. Background agents let it research several tickers at once and pull the results back together.

4:00 午後 - 5:00 午後 (UTC)

The last episode covers what you need before any of this runs somewhere real. OpenTelemetry traces and logs show exactly what the agent did on each turn. Purview handles governance and data protection. Evaluation gives us a way to tell whether quality holds as the SDK keeps moving. Then we deploy to Foundry Hosted Agents so the thing runs as a service instead of a console app. We close by walking through the full refresh-prompt workflow as a maintenance pattern worth stealing.