# Session 1 Handout: Why Agents Fail and How to See It

<p class="meta">FAST-13 Evaluating and Operating Agentic AI Systems · 7formations · Instructor: Wadood Chaudhary · All examples fictional</p>

## Why this session matters

Agents pick tools, pass arguments, loop, and act. A wrong step early (a swapped digit in an ID) can produce a fluent, confident, wrong answer. You can't fix what you haven't named, so the course starts with reading traces and naming failures.

## Key concepts

### 1. Agent anatomy

| Part | What to evaluate |
| --- | --- |
| Router | Did it pick the right tool, or correctly pick none? |
| Tool call | Were the arguments right? |
| Skill | Retrieval, SQL, or code quality |
| Trajectory | The sequence of steps |
| Outcome | The end state (a row written, a ticket opened) |

### 2. Vocabulary

- **Task:** one test case with inputs and success criteria.
- **Trial:** one attempt at a task. Agents vary, so run several.
- **Grader:** logic that scores part of a trial (code, LLM judge, or human).
- **Trace (transcript):** the full record of a trial: messages, tool calls, results.
- **Outcome:** the final state of the environment, which can differ from what the reply claims.

**Tests vs. evals:** unit tests check deterministic code. Evals measure model behavior as rates across many tasks and trials. You need both.

### 3. Error analysis

1. **Open coding:** for each trace, note the *first* thing that went wrong, as a cause ("digits swapped in shipment ID"), not a symptom ("bad answer").
2. **Axial coding:** group notes into 4 to 7 failure modes, each with a one-sentence definition.
3. **Count and rate severity**, then pick the top modes by count × severity.
4. **Propose a grader** for each: input and pass condition. Prefer code where possible.

## Worked example (fictional ShipTrack traces)

| Mode | Count / 40 | Severity | First grader |
| --- | --- | --- | --- |
| Wrong ID extracted | 7 | High | Code: regex ID from user message = tool argument |
| Invented ETA (no get_eta call) | 6 | High | Code: ETA in reply requires a get_eta span |
| Handoff loop | 4 | Medium | Code: no assistant reply after open_handoff |
| Stale status shown as current | 3 | Medium | LLM judge with tool result |
| Off-topic answered | 3 | Low | LLM judge or keyword rule |
| OK | 17 | | |

## Checklist

- [ ] 30+ traces open-coded with causes, not symptoms.
- [ ] 4 to 7 modes, each with a definition someone else could apply.
- [ ] Counts and severity recorded.
- [ ] Three graders proposed, at least one code-based.
- [ ] W1 submitted.