Part 3 · Working With AI Tools
Agentic AI Workflows
In the last lesson you built a single skill and gave it one clear instruction to carry out. Now picture typing one line: pull last quarter's support tickets, group them by theme, draft a summary. Then you walk away. Fifteen minutes later it is done. Not a reply. A finished piece of work.
Why This Matters
Until now, you have prompted AI one turn at a time. Agentic tools change the unit of work: you hand over an outcome, and the tool plans and runs the steps to reach it. That is where real productivity lives, and real risk. By the end of this lesson you will:
- Delegate a multi-step task to an agentic tool by specifying the outcome
- Set boundaries that keep it inside a safe scope
- Evaluate the result before you accept it
This is the core of Specification Thinking and orchestration.
What makes a workflow agentic
A chatbot answers a question. An agent completes a task. Give an agentic tool a goal and it plans a sequence of steps, uses tools to carry them out, and hands back a finished result. Claude Cowork, Gemini Spark, ChatGPT agent mode and Microsoft Copilot agents all work this way.
The right mental model is the one from the last lesson: an agent is a scoped task you start and stop, and you govern it like a tool. Treating it as a member of staff is a category error.
Workplaces are testing that idea. Microsoft's Entra gives an agent its own sign-in identity the moment it is created, and some vendors market them as digital employees with names and shifts. The identity itself argues for the tool model: it exists to audit what the agent did, to limit what it may touch, and to record the human who owns it. The identity is governance plumbing; it makes the agent answerable.
Multi-step autonomy is exactly where the risk sits. A chatbot's mistake is one wrong answer you can see. An agent's wrong assumption at step one flows silently into every step after it, so the error arrives buried inside a polished-looking result. Delegation is therefore not fire-and-forget.
Three moves make it safe. Specify the outcome, so the tool aims at what "done" means. Set the boundaries: which files, systems and accounts it may touch, and how far to go. And define the review gate before you start: decide what you will check before you accept anything. Human-in-the-loop is a gate you set up front, before the work runs.
- Specify the outcome: define what done means, not just the topic.
- Set the boundaries: name what it may touch, and how far to go.
- Define the review gate: decide your checks before you start.
Key Takeaway
One habit makes delegation safe: write your review gate before you delegate, never after. Skip it and you end up trusting a result you never checked. Watch for scope creep. An agent doing more than you asked is the first sign your boundaries were too loose.
Check your understanding
Self-check only. Nothing is scored and nothing is recorded; open an answer to compare it with your own.
Aisha manages partnerships for a non-profit. Every Monday she manually checks 5 partner websites for news updates, compiles a summary, and emails it to her director. She decides to delegate this to an AI agent. She writes: 'Check these 5 partner websites for news from the last 7 days, summarise each update in 2 sentences, and compile into a formatted email draft.' The agent produces a clean summary email, but one 'update' is actually a press release from 3 months ago that was recently reshared on the partner's social media. What boundary was missing from Aisha's delegation brief?
- She should have specified which AI platform to use, some are better at date filtering than others.
- She should have told the agent to only include items published in the last 7 days based on the article's publication date, not the date it appeared on social media or was reshared, and to flag any items where the publication date is unclear.
- She should have asked the agent to include links to each source so she can verify dates manually.
- She should run the agent twice and compare results to catch inconsistencies.
Show the answer
She should have told the agent to only include items published in the last 7 days based on the article's publication date, not the date it appeared on social media or was reshared, and to flag any items where the publication date is unclear.
Tom uses Gemini Spark to manage his calendar and inbox. Spark sends him a Daily Brief each morning with priorities, meeting prep summaries, and suggested email responses. One morning, Spark's brief recommends declining a meeting with a key client because 'your calendar shows three back-to-back meetings that day, and your previous pattern suggests you cancel the least strategic meeting when overbooked.' Tom's relationship with this client is the most important one this quarter. What does this situation illustrate about agentic AI oversight?
- Gemini Spark is not sophisticated enough for calendar management. Tom should switch to a different agent.
- Tom should turn off Spark's recommendation features and only use it for summaries, not decisions.
- Agents can identify patterns but not priorities. Tom should set explicit boundaries, flag key clients as 'never suggest declining' and mark strategic priorities that override scheduling patterns.
- This shows that agentic AI should not have access to calendars, the risk of a bad recommendation is too high.
Show the answer
Agents can identify patterns but not priorities. Tom should set explicit boundaries, flag key clients as 'never suggest declining' and mark strategic priorities that override scheduling patterns.
Saved in this browser only. Carry progress between devices with the resume link on the course page.