Agent Evaluations
Synopsis
Agent evaluations are the measurement layer for systems that plan, call tools, write code, retrieve context, or take actions over time. They combine offline tests, production traces, human review, model-as-judge scoring, regression suites, and task-specific rubrics so teams can tell whether an agent is actually improving rather than merely sounding better.
Origin And Context
The practice grows out of software testing, information-retrieval benchmarks, ML evaluation, and LLM prompt evaluation. Agentic systems made the problem harder because success depends on multi-step behavior: tool choice, state handling, recovery, cost, latency, safety, and final task outcome.
Why It Matters
Without evaluations, agent teams cannot safely change prompts, models, tools, routing, memory policies, or autonomy levels. Evals turn vague quality complaints into visible failure modes and make it possible to ship agents with rollback criteria, measurable acceptance thresholds, and a shared language for product and engineering decisions.
How To Use It
Start with real traces and representative tasks. Define the outcome that matters, add rubrics for intermediate behavior, keep golden examples for regressions, and separate fast pre-merge checks from slower production audits. Use model judges only when their decisions are calibrated against human review, and track cost, latency, and failure categories alongside quality.
Where It Is Useful
Evaluations are useful in coding agents, support agents, research agents, data agents, voice agents, retrieval systems, and any workflow where the agent can take a plausible but wrong path. They are especially valuable where correctness, trust, or operational cost matters.
When To Use It
Use evals before launching, whenever prompts or models change, when adding new tools, after incidents, and when expanding an agent into a new user segment or task family. Lightweight evals should run continuously; deeper reviews should run before major releases.
Active Use Cases
- Regression tests for prompt, model, and tool changes.
- Production trace review for agent reliability and cost drift.
- Benchmarking coding agents, retrieval agents, and long-horizon workflows.
- Reward-signal generation for continual learning and fine-tuning loops.
Related Slide Decks
- youtube aHhB3sjGjkI slides — Agents Building Agents - Alfonso Graziano, Nearform (24 extracted slide frames)
Related Scheduled Sessions
- 2026 06 30 maor bril evaling video slop — Evaling Video Slop; Maor Bril (Day 3 — Session Day 2 · 1:55pm-2:15pm · Evals; official schedule)
- 2026 06 29 laurie voss from vibes to production evaluating and shipping ai agents that work 101 — From Vibes to Production: Evaluating and Shipping AI Agents That Work 101; Laurie Voss (Day 1 — Workshop Day · 9:00am-11:00am · Track 1; official schedule)
- 2026 06 29 laurie voss from vibes to production evaluating and shipping ai agents that work 201 — From Vibes to Production: Evaluating and Shipping AI Agents That Work 201; Laurie Voss (Day 1 — Workshop Day · 2:20pm-4:20pm · Track 1; official schedule)
- 2026 06 30 chris souza model whisperers how evals and prompts shape agent behavior — Model Whisperers: How Evals and Prompts Shape Agent Behavior; Chris Souza, Preetika Bhateja, Daniel Bump (Day 3 — Session Day 2 · 1:30pm-1:50pm · Evals; official schedule)
- 2026 06 29 tejas kumar evals in ai a deep dive — Evals in AI: A Deep Dive; Tejas Kumar (Day 1 — Workshop Day · 12:10pm-1:10pm · Workshops Day 1; official schedule)
- 2026 06 29 wolfram ravenwolf from zero to leaderboard building an end to end ai agent evaluation pipeline — From Zero to Leaderboard: Building an End-to-End AI Agent Evaluation Pipeline; Wolfram Ravenwolf (Day 1 — Workshop Day · 12:10pm-1:10pm · Workshops Day 1; official schedule)
- 2026 06 30 soumya gupta building closed loop evals for a multimodal agent at uber scale — Building Closed-Loop Evals for a Multimodal Agent at Uber Scale; Soumya Gupta, Jai Chopra (Day 3 — Session Day 2 · 11:40am-12:00pm · Evals; official schedule)
- 2026 06 30 rustem feyzkhanov from agent traces to agent simulations the next era of agent evaluation — From Agent Traces to Agent Simulations: The next era of agent evaluation; Rustem Feyzkhanov (Day 3 — Session Day 2 · 12:05pm-12:25pm · Evals; official schedule)
- 2026 06 30 akele reed evals driven development engineering a mental health ai coach ethically and safely — Evals Driven-Development: Engineering a Mental Health AI Coach Ethically & Safely; Akele Reed, Dave Revere, Doug Keller (Day 3 — Session Day 2 · 2:50pm-3:10pm · Evals; official schedule)
- 2026 06 29 nachiket paranjape ai evals platform for cross functional teams at scale — AI Evals Platform for Cross-Functional Teams at Scale; Nachiket Paranjape, Swaroop Chitlur Haridas (Day 2 — Session Day 1 · 1:55pm-2:15pm · AI-Native Enterprises; official schedule)
- 2026 06 29 ari morcos data quality is the compute multiplier — Data Quality is the Compute Multiplier; Ari Morcos (Day 2 — Session Day 1 · 10:45am-11:05am · Data Quality; official schedule)
- 2026 06 30 laurie voss evals track intro — Evals Track Intro; Laurie Voss, Aparna Dhinakaran (Day 3 — Session Day 2 · 10:25am-10:30am · Autoresearch; related YouTube resource; via youtube Xfl50508LZM)
- 2026 06 30 parth asawa beyond static intelligence evaluating continual learning — Beyond Static Intelligence: Evaluating Continual Learning; Parth Asawa (Day 3 — Session Day 2 · 10:45am-11:05am · Memory & Continual Learning; official schedule)
- 2026 06 30 philipp schmid don t ship skills without evals — Don't Ship Skills Without Evals; Philipp Schmid (Day 3 — Session Day 2 · 3:20pm-3:40pm · Evals; official schedule)
- 2026 06 30 laurie voss the death of the code review — The Death of the Code Review; Laurie Voss (Day 3 — Session Day 2 · 12:05pm-12:25pm · AI Architects: Tokenmaxxing; related YouTube resource; via youtube Xfl50508LZM)
- 2026 06 30 laurie voss how long can your skills be before your agent forgets what you told it — How long can your skills be before your agent forgets what you told it?; Laurie Voss (Day 3 — Session Day 2 · 1:30pm-1:50pm · Context Engineering; related YouTube resource; via youtube Xfl50508LZM)
- 2026 06 29 ameya bhatawdekar your agent evolved your evals didn t — Your Agent Evolved. Your Evals Didn't.; Ameya Bhatawdekar (Day 2 — Session Day 1 · 11:10am-11:30am · AI Architects: Show my Workflow; official schedule)
- 2026 06 30 lukas petersson vending bench long horizon agent evals for a simulated vending business — Vending-Bench: Long-Horizon Agent Evals for a Simulated Vending Business; Lukas Petersson (Day 3 — Session Day 2 · 10:45am-11:05am · Evals; official schedule)
- 2026 07 01 ashok chandrasekar are llm performance benchmarks reliable — Are LLM Performance Benchmarks Reliable?; Ashok Chandrasekar, Jason Kramberger (Day 4 — Session Day 3 · 11:40am-12:00pm · Inference; official schedule)
- 2026 07 01 session vector isn t enough hybrid search and retrieval for ai engineers — Vector Isn't Enough: Hybrid Search & Retrieval for AI Engineers; Jeff Vestal (Day 1 — Workshop Day · 2:20pm-4:20pm · Track 7; official schedule)
- 2026 06 29 jess wang agentic vs vector search an eval driven approach to coding agent performance — Agentic vs. Vector Search: An Eval-Driven Approach to Coding Agent Performance; Jess Wang (Day 2 — Session Day 1 · 11:40am-12:00pm · Expo Stage 2 NW; official schedule)
- 2026 06 29 simran arora can llms write fast multi gpu kernels we built a benchmark to find out — Can LLMs write fast multi-GPU kernels? We built a benchmark to find out.; Simran Arora (Day 2 — Session Day 1 · 12:05pm-12:25pm · Expo Stage 3 SW; official schedule)
- 2026 06 30 ali khial benchmarks the good the bad and the ugly — Benchmarks: The Good, the Bad, and the Ugly; Ali Khial (Day 3 — Session Day 2 · 3:20pm-3:40pm · Posttraining & Midtraining; official schedule)
- 2026 06 29 doug guthrie advanced workshop mastering ai observability — Advanced workshop: Mastering AI Observability; Doug Guthrie (Day 1 — Workshop Day · 9:00am-11:00am · Track 9; related YouTube resource; via youtube bk0TmxoZlUY)
Related People
- Laurie Voss
- Pamela Fox
- Fuad Ali
- Brendan Rappazzo
- Ahmad Osman
- Yuval Belfer
- Harshul Jain
- Tanmay Sah
- Filip Makraduli
- Frank Coyle
- Maor Bril
- Chris Souza
- Preetika Bhateja
- Daniel Bump
- Tejas Kumar
- Wolfram Ravenwolf
- Soumya Gupta
- Jai Chopra
- Rustem Feyzkhanov
- Akele Reed
- Dave Revere
- Doug Keller
- Nachiket Paranjape
- Swaroop Chitlur Haridas
Related Companies
- Arize AI
- Microsoft
- Uber
- Braintrust
- Meta
- Towards AI
- Weights & Biases by CoreWeave
- NVIDIA
- AI21
- SonderMind
- DoorDash
- Elastic
- Digital Ocean
- Arize
- turbopuffer
- poolside
- G2i
Transcript And Resource Support
Transcript-backed resources
- youtube Xfl50508LZM — Ship Real Agents: Hands-On Evals for Agentic Applications — Laurie Voss, Arize
- youtube bk0TmxoZlUY — Evals 101 — Doug Guthrie, Braintrust
- youtube iNkFlCiij0U — The Art & Science of Benchmarking Agents — Vincent Chen, Snorkel AI
- youtube vljxQZfJ9wY — Production Evals For Agentic AI Systems - Nishant Gupta, Meta Superintelligence Labs
- youtube pSto5YaNGUo — The Agentic AI Engineer - Benedikt Sanftl, Mutagent
- youtube hqHC6Z_lXyo — 20 days of compute vs 7 hours: rethinking what state-of-the-art means — Bertrand Charpentier, Pruna
- youtube YYH0DMQr30A — Task Fidelity Scaling Laws — Kobie Crawdord, Snorkel
- youtube aHhB3sjGjkI — Agents Building Agents - Alfonso Graziano, Nearform
- youtube 2IxD9OB3XuQ — Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI
- youtube QuuIywMG4s8 — Evals Are Broken, Use Them Anyway — Ara Khan, Cline
- youtube ObTPqBGsEbA — The Production AI Playbook: Deploying Agents at Enterprise Scale — Sandipan Bhaumik, Databricks
- youtube T0HhO4YtTfE — AI System Design: From Idea to Production - Apoorva Joshi, MongoDB
- youtube htM02KMNZnk — WF2026: Software Factories & Keynotes ft. Microsoft, OpenAI, OpenClaw, Z.ai (GLM), MiniMax, HF
- youtube Jx4ZFEAq6bY — User Signal Dies at the Retrieval Boundary - Sonam Pankaj, StarlightSearch
- youtube wcUJWP6WpGM — SWE-rebench: Lessons from Evaluating Coding Agents — Ibragim Badertdinov, Nebius
- youtube LrGCT7G_rU8 — Using RL Agent to Detect and Remediate ETL Pipeline Failures - Anna Marie Benzon
- youtube IJXjTLPzvAU — The Miranda Hypothesis: How Hamilton Poisoned Persona Evals - Jacob E. Thomas, Results Gen
- youtube TNwJ1LMiENk — Stop Making Models Bigger, Make Them Behave — Kobie Crawford, Snorkel
Quote signals
- “Uh I had a question on um on how how much evaluation you need to write for uh feature cuz especially when you run against live traces, sometimes the the evaluation can cost more than the actual feature.” — youtube Xfl50508LZM
- “One of the key things is it treat memory as reasoning, not as facts, statistics, fact with no context and no history, but reasoning.” — youtube Jx4ZFEAq6bY
- “But the problem is that state of the art is a bit a confusing concept and people maybe have different vision on this.” — youtube hqHC6Z_lXyo
- “The event outcome becomes a first-class signal in the retrieval re-ranking and not just for retrieval.” — youtube Jx4ZFEAq6bY
- “Uh the other thing to call out here is like we already have a lot of companies using brain trust in production today.” — youtube bk0TmxoZlUY
- “Um so anyway, so the the title of my talk today is evals are broken and you should use them anyway.” — youtube QuuIywMG4s8
- “And lastly, this axis is all about producing more complex work, more representative work, and also nuanced signals that can be used for not just evaluation, but reward signals during training.” — youtube iNkFlCiij0U
- “And for the same amount of evaluation, it takes only 7 hours.” — youtube hqHC6Z_lXyo