Module 3 — Tool Calling and Execution

A language model cannot read a file, run a test, or delete a branch. Everything an agent does it does through a tool call, and every tool call is a round trip through a contract the harness defines on both ends: the harness tells the model what exists and how to ask, the model emits a request in a format it has only statistically internalized, and the harness must dispatch that request safely, bound the result, and hand back something the model can act on. Module 2 built the loop that keeps this cycle turning. This module builds what the cycle turns on. It is the largest module in the course because tool design is where harness quality is actually decided: two harnesses running the same model on the same task diverge almost entirely on the quality of their tool surface — the schemas, the error messages, the truncation strategy, the concurrency rules. A model with excellent tools and a mediocre loop outperforms a model with an excellent loop and vague, unbounded, silently-failing tools every time.

Learning objectives

By the end of this module you can:

• M3.1 Explain the tool-calling contract end to end: schema advertisement, model emission, harness dispatch, result return.
• M3.2 Design tool schemas that models actually use correctly — naming, description economy, parameter shape, enum over freeform, required vs optional.
• M3.3 Validate and coerce arguments defensively; fail with messages that teach the model to retry correctly.
• M3.4 Walk the tool execution pipeline: resolution, permission check, execution, result shaping, error surfacing.
• M3.5 Shape tool results for a token budget — truncation, pagination, structured summaries, and the line-number/offset idiom.
• M3.6 Handle parallel and concurrent tool calls, including ordering and interference hazards.
• M3.7 Implement the core tool families: filesystem, shell/subprocess, search, web fetch, and structured editing.
• M3.8 Compare tool interop standards — MCP, ACP, and native registries — and when to adopt each.
• M3.9 Test tools: unit, snapshot, and real-API tiers.

Prerequisites and time

Module 1 (the ~150-line Module 1 spine) and Module 2 (the agent loop, phases, cancellation, and the tools/execute waterfall you met through the timeout policy). You should be comfortable reading TypeScript interfaces and JSON Schema.

Estimated time: 6 instructional hours plus a 2-hour lab. This domain is 20% of the certification exam — 13 of 65 items.

---