Agent Design, Optimization and Harnesses
Choose a control policy, improve it on separated task data and independently verify bounded local actions.
- Prerequisites
- A prior ADK or LangGraph course, tool contracts, retrieval and LLM Evals. These modules use those components rather than teach their introductions again.
- Duration
- about 4 hours of reading
- Cost
- Free, no sign-up
What you can build now
The design and optimization modules use the shared docuctl task lab. The harness module moves from controlled expression repairs to a trusted repository-tool replay, then applies independent outcome checking to controlled candidates. These are linked experiments with different task contracts, not one deployed coding assistant. Persistent memory lifecycle belongs to Context and Memory.
What you'll learn
- Compare workflows and bounded policies using the same task cases.
- Curate traces, tune on development tasks and calibrate independent graders.
- Run bounded local repair and coding-agent loops, then verify edits independently.
Lessons
Module 1 · 3 lessonsAgent design
Compare a workflow with a bounded policyRead
Compare a keyword route, fixed search/read workflow and scripted decision loop on shared documentation tasks. Measure outcomes and tool work.
Run a search-and-read decision loopRead
Separate policy decisions from tool execution. Enforce action schemas, tool and turn limits, and citation provenance before checking the returned command.
Delegate independent tasks with explicit worker stateRead
Fan out two documentation requests into separate bounded runtimes. Gather results by task ID and account for dispatch, merge and incomplete-worker behavior.
Module 2 · 3 lessonsAgent optimization
Curate traces before selecting an improvementRead
Deduplicate versioned questions, keep task families within one split and save the actual decisions and tool observations for each cleaned case.
Tune retrieval policy on separated task familiesRead
Choose alias expansion and candidate count on development tasks. Freeze the selected configuration before measuring source recall on separate task families.
Calibrate task graders before comparing policiesRead
Expose false positives in a format-only grader, compare it with the task contract and calculate what any-success and repeated-success metrics actually mean.
Module 3 · 3 lessonsAgent harnesses
Bound a local repair processRead
Restrict a retry-condition patch, write each candidate into a fresh workspace and run independent behavioral checks in a separate process with a timeout.
Build a coding-agent loopRead
Build a coding-agent loop that streams model messages, runs repository tools, checks a real test repair and stops at a fixed turn limit.
Verify edits independently of completion claimsRead
Run controlled repair candidates through the existing harness. Preserve completion claims as observations while grading the behavior of each fresh candidate.