MCP and agent evaluation: finish Project 3.

1. 1.Try #19 Evaluate Reverse Polish Notation – untimed, AI-free

   Stack · Medium. Aim for 15–20 minutes. No assistants, no autocomplete.

   Open on LeetCode (opens in a new tab)
2. 2.Stuck at 25 minutes? Use the NeetCode method

   Read or watch the explanation, close it, then re-code the solution from memory.

   NeetCode 150 explanations (opens in a new tab)
3. 3.Write the Big-O

   Add a comment with the time and space complexity, and why.

4. 4.Log it

   Mark it Solved or Revisit, with minutes and whether you solved it unaided.

   Log #19 on the Problems page
5. 5.10 min revision

   One item from reviews/relearn.md.

Companion video

The Model Context Protocol (MCP)Anthropic · Mon–Wed

1. 1.Try #20 Daily Temperatures – untimed, AI-free

   Stack · Medium. Aim for 15–20 minutes. No assistants, no autocomplete.

   Open on LeetCode (opens in a new tab)
2. 2.Stuck at 25 minutes? Use the NeetCode method

   Read or watch the explanation, close it, then re-code the solution from memory.

   NeetCode 150 explanations (opens in a new tab)
3. 3.Write the Big-O

   Add a comment with the time and space complexity, and why.

4. 4.Log it

   Mark it Solved or Revisit, with minutes and whether you solved it unaided.

   Log #20 on the Problems page
5. 5.10 min revision

   One item from reviews/relearn.md.

1. 1.Try #21 Car Fleet – untimed, AI-free

   Stack · Medium. Aim for 15–20 minutes. No assistants, no autocomplete.

   Open on LeetCode (opens in a new tab)
2. 2.Stuck at 25 minutes? Use the NeetCode method

   Read or watch the explanation, close it, then re-code the solution from memory.

   NeetCode 150 explanations (opens in a new tab)
3. 3.Write the Big-O

   Add a comment with the time and space complexity, and why.

4. 4.Log it

   Mark it Solved or Revisit, with minutes and whether you solved it unaided.

   Log #21 on the Problems page
5. 5.10 min revision

   One item from reviews/relearn.md.

1. 1.Work through: DeepLearning.AI – MCP: Build Rich-Context AI Apps with Anthropic

   lessons up to connecting a server

   Open (opens in a new tab)
2. 2.Do

   Follow the lessons.

3. 3.Check you’re done

   Done when: The course server runs.

1. 1.Do

   Wrap your 2 tools as an MCP server and connect a client.

2. 2.Check you’re done

   Done when: The client lists and calls your tools.

1. 1.Catch up first

   Finish anything unticked from Mon–Fri this week before stretch work.

   Week 12 sessions
2. 2.Then stretch work

   DeepLearning.AI Evaluating AI Agents: remaining lessons.

3. 3.Level up: Prompt caching

   The provider stores the unchanging start of your prompt (system prompt, tool definitions, long documents) and bills re-reads at a fraction of the normal input price. Claude’s docs list cache reads at 0.1× the base input price, with 5-minute or 1-hour lifetimes. Turn on caching for P3’s system prompt and tool definitions. Print cache-read and cache-write tokens from the API response for a 5-step run, before and after. Done when: P3’s README shows cost per task with and without caching.

   Prompt caching (opens in a new tab)

1. 1.Start with 20 min revision

   Explain agent vs workflow.

2. 2.Do

   Add phs_get_rows and summarise_table; a 25-task eval (right tool? right arguments? answer equals pandas ground truth? steps? cost?); fix 2 failures.

3. 3.Check you’re done

   Done when: ≥80% of tasks pass.

4. 4.Commit

   P3 evals/agent_eval.py

5. 5.Level up: Designing tools for agents

   Tools are a contract between the model and your code. Good ones are named and described for a model rather than a developer, return only what the model needs, and give error messages the model can act on. Take the P3 tool the agent misuses most. Rewrite its name, description and error messages using the article’s advice, change nothing else, and rerun the eval. Done when: The eval table has a before/after row for that one tool-description change.

   Writing effective tools for agents – with agents (opens in a new tab)

Companion video

Building Agents with Model Context Protocol – Full Workshop with Mahesh Murag of AnthropicAI Engineer · Sat

1. 1.Do

   README with a Mermaid diagram, eval table and GIF; weekly review.

2. 2.Check you’re done

   Done when: README committed.

3. 3.Commit

   P3 v1.0

4. 4.Weekly review (last 30 min)

   3 wins, 3 confusions, 1 thing to re-learn – written at the bottom of this page and saved with your progress.

   Open the weekly review
5. 5.Level up: Evals for agents

   An agent run is non-deterministic, so one pass proves little. The vocabulary to know: task, trial, grader, transcript. You evaluate the model and its harness together and run each task several times. Run each of your 25 P3 tasks three times. Report how many pass at least once and how many pass all three times. Done when: The P3 README reports both numbers and names the flakiest task.

   Demystifying evals for AI agents (opens in a new tab)