Skip to content

Agent Evaluation Guide

Evaluating AI agent systems in Potato: coding agents, web agents, RAG pipelines, and multi-agent systems.

Overview

Potato supports evaluating AI agents at multiple levels:

Level What You Annotate Example
Trajectory Overall task success "Did the agent complete the task?"
Step Individual action correctness Per-turn Likert ratings on each agent step
Span Specific text segments Highlight hallucinated claims, factual errors
Comparison Side-by-side evaluation "Which agent performed better?"

Trace Conversion

Import traces from any major agent framework:

python -m potato.trace_converter --input traces.json --input-format openai --output data.jsonl

Supported formats: OpenAI, Anthropic/Claude, ReAct, LangChain, LangFuse, WebArena, SWE-bench, OpenTelemetry, CrewAI/AutoGen/LangGraph, MCP, Aider, Claude Code, ATIF, SWE-Agent, and Web Agent.

For full details, see Agent Traces.

Coding Agent Evaluation

Evaluate agentic coding systems (Claude Code, SWE-Agent, Aider) with:

  • Diff rendering for code changes
  • Process Reward Model (PRM) annotation
  • Code review workflows

See Coding Agent Annotation.

Web Agent Evaluation

Review GUI agent traces with an interactive screenshot viewer:

  • Step-by-step navigation through screenshots
  • SVG overlays showing clicks, bounding boxes, mouse paths, and scroll actions
  • Inline annotation controls per step
  • Live browsing mode with automatic trace recording

See Web Agent Annotation.

Trajectory Evaluation

Per-step error marking with typed error taxonomies:

  • Mark individual steps as correct, incorrect, or partially correct
  • Assign error types (hallucination, reasoning error, tool misuse, etc.)
  • Span-level annotation within agent output

See Trajectory Evaluation.

Live Agent Interaction

Observe and interact with a live AI agent in real time, recording traces as you go:

See Live Agent Interaction.

Continuous Evaluation Loop

Beyond annotating individual traces, Potato closes the loop from production traffic to graded, regression-gated evaluation:

  • Capture — instrument your agent with the @traceable tracing SDK (potato_trace), or POST traces to the ingestion webhook / Langfuse poller.
  • Automate — automation rules run filter → sampling rate → actions over every incoming trace, routing it to the annotation queue, an eval dataset, an evaluator, an outbound webhook, or a notification.
  • Curate — build versioned datasets & experiments (file or SQLite, tagged versions, splits), and use the semantic Catalog (embedding search + dynamic slices) to find the traces worth reviewing.
  • Evaluate — score outputs with programmatic evaluators (deterministic trajectory match, tool-use correctness, reference-free LLM-judge, heuristics), or compare models head-to-head in the model arena.
  • Gate CI — run evals inside pytest and fail the build on score-threshold regressions with the CI evaluation plugin.
  • Calibrate judges — measure and tune an LLM judge against human gold with judge alignment (Cohen's κ, auto-calibration from human corrections) and blind multi-LLM judge calibration; judges work on categorical, span, and free-text outputs.

Inspect and control all of this from the admin dashboard (Datasets & Experiments, Automation, Catalog, Arena links) and the /admin/eval/* API.

Using AI Assistance for Evaluation

Speed up agent evaluation with AI-powered features:

  • AI Support - LLM label suggestions for agent evaluation tasks
  • Chat Support - Ask an LLM questions about complex agent traces

Example Configurations

Ready-to-use examples in examples/agent-traces/:

Example What It Evaluates
agent-trace-evaluation/ Text agent traces with MAST error taxonomy
visual-agent-evaluation/ GUI agents with screenshot grounding
agent-comparison/ Side-by-side A/B agent comparison
rag-evaluation/ RAG retrieval relevance and citation accuracy
openai-evaluation/ OpenAI Chat API traces with tool calls
anthropic-evaluation/ Claude messages with tool_use blocks
swebench-evaluation/ Coding agents with patch correctness ratings
multi-agent-evaluation/ Multi-agent coordination (CrewAI, AutoGen, LangGraph)
web-agent-review/ Pre-recorded web traces with overlay viewer
web-agent-creation/ Live web browsing with trace recording

Run any example:

python potato/flask_server.py start examples/agent-traces/agent-trace-evaluation/config.yaml -p 8000