Tool of the Week: Copilot Studio's agent node, an AI agent as one step in a workflow
Microsoft's Copilot Studio agent node reaches general availability this month. It lets you call a published agent as a single step inside an automated workflow. The agent can reason over data, pull knowledge and use tools, then hand a result to the next step.
That sounds small. It is the pattern most real automation work needs. A workflow that moves a file or updates a record is easy. The hard step is the one in the middle where someone reads an email, checks it against something and decides. That step is why many processes still depend on a person remembering to do it. An agent node is Microsoft's answer for putting a decision there without writing custom logic.
Two details matter if you run IT or ops. It is enabled by default for makers, which means people in your tenant can start wiring agents into workflows without asking you. And administrators can restrict it with data loss prevention policies. If you own governance, that policy is the control to look at before the feature shows up in someone's workflow.
My take: the build is the easy part. The value is in choosing which step deserves an agent and which should stay a plain rule. Use an agent where the input is messy text and the decision needs judgment. Keep deterministic steps deterministic. Log what the agent decided, because someone will ask later.
Preview started April 1. Availability is worldwide on standard multi-tenant clouds.
Quick Hits
Anthropic cut Claude Opus 5.5 prices about 40 percent below Opus 5. Launched September 22. API pricing is $4 per million input tokens and $20 per million output, down from $5 and $25. Reporting says it matches the work quality of the more expensive Fable 5.1 model, and it is available through the API, AWS, Google Cloud, Azure and the paid Claude plans. If an agent workflow was priced out at the Opus tier, rerun the cost estimate.
OpenAI answered 90 minutes later with GPT-6 Sol and Luna at half the GPT-5.6 price. Sol is $2 input and $10 output per million tokens. Luna, the small one, is $0.10 and $0.50. Both are half the price of their GPT-5.6 counterparts, per The Next Web. For high-volume jobs like classifying inbound files or summarizing tickets, Luna-class pricing changes which processes are worth automating. Test on your own data before you trust a benchmark.
Prompt of the Week: The Worth-Automating Filter
I will describe a recurring process in my week. Do not suggest tools yet.
Process: [describe it in 3 to 5 sentences: who starts it, what
inputs arrive, what steps a person does, what the output is]
Answer these in order:
1. FREQUENCY: how often does this run, and how long does one run take?
Multiply out the hours per month.
2. DECISIONS: list each step where a person uses judgment versus
follows a fixed rule. Mark the judgment steps.
3. FAILURE: what happens if a run is skipped or done wrong, and who
notices first?
4. INPUTS: how messy are the inputs (clean fields, semi-structured
files, free text)?
5. VERDICT: AUTOMATE (fixed rules, plain script or workflow),
AGENT STEP (needs judgment on messy input, human reviews output),
or LEAVE ALONE (rare, low stakes, or not worth the setup).
One sentence on why.
If the hours per month are under 2, say LEAVE ALONE and stop.Run it on three processes before you build anything. The answer is often "a plain script," and that is fine.
One last thing
Writing this is one side of the work. The other is building it: data pipelines, file and reporting workflows, integrations between systems that were never meant to talk, and the automation that takes the manual steps out.
If a process in your week still depends on someone remembering to run it, book a free 15-minute audit and I will find your top 3 time-wasters. If it is not worth automating, I will tell you that.
Scott