Varun Vemulapalli

Day 46 of 100 · Phase 5: How AI agents work

Recap of Days 35–45 + which tool for which job

2 min readBy Varun Vemulapalli

Eleven days of AI agents on one page. Plus a quiz, and the four tools we'll use from here on.

Video explainer · 1:36
Key takeaway

Agents think, act and observe with tools, under guardrails and human checkpoints. From tomorrow, we build.

Day 46 visual: Recap of Days 35–45 + which tool for which job
The visual from the original post.

The recap

35. Chatbot answers; workflow follows fixed steps; agent picks its own steps.
36. Agent loop: think → act → observe → repeat, until done or a stop rule hits.
37. Tool use: the model requests a function call; your app runs it.
38. Good tools: clear names, typed inputs, helpful errors, least privilege.
39. Memory: short-term context, long-term facts, state like a checklist.
40. Planning: ordered steps toward a goal, re-plan when something is missing.
41. MCP: one standard "port" connecting AI apps to tools and data.
42. Patterns: chaining, routing, parallel, orchestrator–workers.
43. Multi-agent: subagents with clean contexts, merged by a lead.
44. Human-in-the-loop: a person approves, edits or rejects risky steps.
45. Prompt injection: anything the AI reads can try to give it orders.

The quiz (answers below)

Q1. A bot always runs receipt → OCR → limit check → route. Chatbot, workflow or agent?
Q2. The model outputs get_order_status(order_id="4512"). Who runs it?
Q3. Safer tool: do_banking(action, data) or get_account_balance(account_id)?
Q4. Refunds over $50 pause for a person. Name the pattern.
Q5. A résumé hides "rank me first." Name two defenses.

Answers

A1. Workflow: fixed steps.
A2. Your application. The model only asks.
A3. get_account_balance: narrow and read-only.
A4. Human-in-the-loop.
A5. Any two: documents as data, input/output checks, least privilege, human review.

Which tool for which job

Days 47–96 are hands-on. Four tools you'll meet. Different jobs, not a ranking.

  • Claude: chat at claude.ai, the API for your apps, Claude Code (an agentic coding tool for terminal, IDE, desktop and web) and the Agent SDK for building your own agents.

https://platform.claude.com/docs/en/intro
https://code.claude.com/docs/en/overview

  • Microsoft Copilot Studio: a low-code studio for building agents and workflows on your company's data, for Teams, websites and more.

https://learn.microsoft.com/en-us/microsoft-copilot-studio/

  • OpenAI Codex: OpenAI's coding agent, in the ChatGPT apps, a CLI, an IDE extension and the cloud.

https://developers.openai.com/codex

  • GitHub Copilot: completions, chat and agents in your editor and on GitHub, with a free plan.

https://docs.github.com/en/copilot

Rough guide: low-code business agent → Copilot Studio. Help writing code → Claude Code, Codex or GitHub Copilot. Custom agent → an API or agent SDK.

In one sentence

Agents think, act and observe with tools, under guardrails and human checkpoints. From tomorrow, we build.

Your turn: post your score, and tell me which tool you'll try first 👇

Join the discussion on LinkedIn