Stanford CS329A Self‑Improving AI Agents ｜ Part 8 ｜ Agentic Evaluations and Long Horizon Tasks — Annotated Lecture Notes & Key Takeaways

Course Information

Institution: Stanford

Original Course: Stanford CS329A Self‑Improving AI Agents ｜ Part 8 ｜ Agentic Evaluations and Long Horizon Tasks — Annotated Lecture Notes & Key Takeaways

Instructor Bio: This session is co-instructed by **Aakanksha Chowdhery** and **Azalia Mirhoseini**, faculty and researchers at Stanford University and Google DeepMind. Aakanksha Chowdhery earned her PhD from Stanford University and is a Senior Researcher at Google DeepMind. She led end-to-end training of the 540B PaLM model and drove pre-training and scaling efforts for multiple generations of Gemini MoE models. She has also contributed core components to PaLM-E, Med-PaLM, and the Pathways infrastructure that underpins Google’s large language model ecosystem, with deep expertise in dense and Mixture-of-Experts architectures and large-scale model training. Azalia Mirhoseini is an Assistant Professor of Computer Science at Stanford University and founder of the Scaling Intelligence Lab. She also serves as a Senior Staff Scientist at Google DeepMind and co-founded Ricursive Intelligence. Previously she held research roles at Google Brain and Anthropic, contributing to the development of the Claude and Gemini model families. Her pioneering work spans MoE architectures, test-time compute scaling, and deep reinforcement learning for chip design (AlphaChip), with a core research focus on recursive self-improving AI systems.

Course Description: This lecture addresses the unique challenges of evaluating self-improving AI agents, particularly on long-horizon tasks where performance unfolds over extended sequences of action and reasoning. It covers agentic evaluation frameworks, benchmark design principles, and metrics for measuring iterative improvement over time. It also examines the limitations of standard static benchmarks and explores new evaluation paradigms that capture adaptive growth, autonomy, and real-world task completion.

These study notes cover Part Eight of Stanford’s CS329A lecture, whose central topic is agentic evaluations. Sitting in on this talk, I walked away thinking most people building agent‑based systems do not put nearly enough thought into proper evaluation work. The session draws heavily on back‑and‑forth discussion between the lecturer and enrolled students. It unpacks three modern benchmarks built to test long‑horizon, real‑world agent performance, moving past older, oversimplified benchmarking approaches.
 

The lecture opens with the instructor polling the room to see what evaluation workflows students actually use for their own class projects. Few hands go up at first. Students bring up practical examples: multi‑hop question‑answering datasets, internal private‑wiki QA workflows, and pipelines feeding wiki notes into coding agents. This opening exchange sets the tone: lots of folks build agent prototypes, but thorough agentic evaluations remain far from standard practice.
 

Moving Past Saturating Traditional Benchmarks

 

Older benchmarks built for chatbot testing, basic question answering and single‑shot prompting are hitting diminishing returns. Just a few years back, the simple fact that a model could generate coherent explanatory text counted as major progress. Today these same tests barely help us tell modern large models apart.
 

The lecturer makes the case that we now need to assess two critical dimensions. One is raw capability: how complex, how lengthy a task can an agent actually finish end‑to‑end? The second is real‑world practical value: can this system deliver work that paid human professionals would normally complete? Two distinct measurement frameworks get introduced here. METR grounds its measurements against time estimates from working domain experts. GDP‑VAL instead calculates win rates by directly comparing agent outputs against finished work from seasoned industry professionals. Both frameworks have valid use cases, yet they paint noticeably different pictures of model progress.
 

We briefly revisit well‑known flaws from early chat systems. Not too long ago, multi‑turn conversations would fall apart quickly, with models losing track of earlier context partway through a dialogue. Contemporary models hold context far more reliably. This capability raises a bigger practical question: how long‑running of a task can an agent sustain, and how consistently can it wrap things up successfully.
 

METR does not just log whether an agent completes a task. It ties each assignment to explicit success thresholds, such as hitting a 50 % or 80 % success rate across repeated runs. Every task in the suite gets anchored to wall‑clock time estimates from human professionals doing that exact same piece of work.
 

Three separate task suites compose the full METR benchmark. The smallest suite covers tiny, atomic actions that take human workers one to thirty seconds — think simple software engineering steps like opening a single file. Edgecast covers assignments ranging from one minute up to thirty hours, spanning general software work and research‑engineering chores. Rebench holds full‑scale machine‑learning research tasks, some taking as long as eight hours for humans to finish. Combined, the benchmark contains roughly 170 unique tasks.
 

Collecting human baseline data for METR brings real‑world headaches. Researchers bring in people with around five years of relevant professional experience to complete tasks, log their timings, and rate subjective task difficulty. Seasoned practitioners often underestimate how tricky a job really is, and their sense of difficulty frequently does not align with where AI agents stumble.
 

One part of this discussion felt a little underdeveloped to me. A student points out human completion timings can vary wildly between different people. Hiring large panels of domain experts costs substantial money, so teams work with smaller groups and measure inter‑rater agreement. If people’s answers diverge significantly, they gather more samples; if most reviewers converge, they stop collecting human data. The lecturer touches on this logic, but does not go deep on where exactly you draw that cutoff line in practice.
 

METR Benchmark: Time‑Horizon Trends, Reliability Gaps and Failure Modes

 

METR plots use human‑estimated task duration along the horizontal axis and agent success rate on the vertical axis. Short, simple atomic tasks see fairly high agent success rates. Mixed‑complexity Edgecast tasks show scattered, inconsistent results. Longer Rebench research‑oriented assignments land at low success rates overall. The broad pattern is intuitive: as tasks grow longer, agents become far more prone to losing their bearings and failing.
 

Looking across successive model releases reveals an eye‑catching pattern. The maximum time‑horizon an agent can handle at a 50 % success rate roughly doubles every seven months. GPT‑2 back in 2019 could only manage work equivalent to two‑second human tasks. GPT‑4 pushed that threshold out to several minutes. By 2025, Claude 3.7 Sonnet reached a 50 %‑success time‑horizon near 59 minutes.
 

That 50 % success number demands careful interpretation. Half your attempts work out, half do not. The lecturer uses a relatable analogy: this resembles handing work over to an intern who only delivers acceptable results half the time. You cannot safely count on consistent, usable output.
 

The performance gap becomes even more obvious once you raise your bar for reliability. At an 80 % success requirement, even top‑tier models can only reliably sustain tasks running about 8‑10 minutes. There is a massive disconnect between what agents can pull off half the time versus what they can deliver most of the time. This gap makes clear just how much room still exists to improve agent reliability.
 

Multiple factors drive better agent performance. Improved logical reasoning, stronger code generation skills, and more capable tool‑use all move the needle. Modern agents also do a better job avoiding pointless repeated actions and recovering when individual steps go wrong. Long assignments will always contain mis‑steps, so agents need to keep their final objective clearly in view and track how much ground they have covered.
 

Students in the room share observations from their own hands‑on work with coding agents. Better context structuring and context compaction logic help a great deal. User‑feedback‑driven online reinforcement learning also moves performance forward. Explicit replanning after collecting new information helps agents reset when things go sideways. Persistent memory systems that build up working knowledge of a codebase also support sustained work across longer workflows.
 

The lecture walks through common agent failure patterns uncovered in METR testing. Agents draft poor high‑level plans and pick inappropriate tools for the job. They make factual or arithmetic mistakes. Sometimes they abandon tasks too early, trapped in repetitive loops where they rerun identical ineffective actions even after those actions have already failed. Older models such as GPT‑4 showed very high loop‑failure rates; reasoning‑optimized models like o‑1 cut down this specific failure mode, yet other error categories persist.
 

No benchmark is flawless, and METR carries clear limitations. It struggles with messy real‑world assignments that lack one single correct answer. Related benchmarks such as SWE‑Bench have their own confounding variables. Many GitHub repositories inside SWE‑Bench already exist inside model training datasets, which artificially inflates reported performance numbers.
 

One observation stuck with me from this segment. When looking at internal company pull‑request work, external contractors unfamiliar with a given repository work 5‑18 times slower than long‑term code maintainers. Agent performance lands much closer to the outside contractor than the deeply knowledgeable maintainer. Agents lack that rich, project‑specific contextual awareness built up over years working on a system. They can contribute meaningfully, yet they operate as outsiders to the project context.
 

GDP‑VAL: Measuring Economic Value Against Human Experts

 

GDP‑VAL takes a completely different approach to evaluation. Instead of focusing on task runtime, it asks a straightforward practical question. Given real‑world work completed by people with ten or more years of professional experience, how frequently can an agent produce output competitive with that expert work? This benchmark draws economically meaningful tasks across nine industry sectors: real estate, government, manufacturing, healthcare, finance, retail, wholesale, media production and others.
 

The example tasks shared feel very concrete. A manufacturing engineer designs a 3D model for an assembly‑line cable‑reel stand. A financial‑services analyst builds a competitive‑landscape research brief. A registered nurse reviews image evidence and writes up consultation notes. A video editor produces an opening reel using a provided script. An auditor hunts for pricing inconsistencies scattered across multiple purchase orders. Many tasks are multimodal, requiring work with spreadsheets, CAD files, audio and video assets.
 

The dataset includes 44 distinct occupations and roughly 320 total tasks. Around 220 of those form the public gold‑standard evaluation split. Most assignments are digital, computer‑based work. Average human completion time runs around seven hours, though some assignments would take human workers weeks to finish. Roughly 70 % of tasks require working with attached reference files. Experts vetted nearly all tasks to ensure instructions were clear and unambiguous.
 

Evaluation uses blind pairwise comparison. Human raters assess outputs without knowing which piece came from a professional and which came from an AI agent. Metrics track two outcomes: pure agent wins, and wins plus ties. Performance trends observed here differ sharply from METR’s exponential time‑horizon growth. Agent win rates improve in a roughly linear fashion. GPT‑4o sat near a 12.4 % win‑plus‑tie rate, while Claude Opus 4.1 reached approximately 47.6 %.
 

This contrast delivers one of the lecture’s most important insights. METR shows agent feasible task length doubling roughly every seven months. GDP‑VAL results suggest real‑world, expert‑comparable performance creeps upward far more slowly. The lecturer warns listeners not to get overly optimistic purely from METR’s exponential curves. Being technically capable of stepping through a multi‑hour task does not guarantee your final output will stand up against work from a seasoned professional.
 

Different model families show distinct strengths under GDP‑VAL testing. Certain multimodal models handle visual aesthetics, document formatting, PDF and spreadsheet parsing particularly well. GPT‑5 fares better with strict instruction adherence, numerical work and text‑heavy assignments.
 

Instruction‑following mistakes stand out as a top failure source. Agents will state they intend to read attached reference material, then completely skip those files and rely on made‑up information instead. Formatting errors also appear regularly. Roughly half of generated outputs count as acceptable yet distinctly sub‑par. Only about 20 % of outputs genuinely outperform human expert work. Close to 30 % of outputs are poor or entirely unusable. Human raters do not always agree on quality judgements, which injects some uncertainty into these statistics.
 

Running multiple attempts combined with self‑revision yields tangible improvements. Parallel or sequential sampling, where the agent critiques and revises its own prior output, delivers around 1.6 % cost‑related improvement and 1.4 % speed‑up relative to human experts for jobs the agent can complete successfully. Even with these gains, agent‑side costs still sit below 10 % of an expert‑level human salary for tasks the agent handles well.
 

Performance varies widely across occupations. Agents reach near‑human parity on select roles including inventory clerks, purchasing agents, some IT‑system managers, and portions of standard software‑engineering workflows. Administrative‑service managers, compliance officers, certain health‑service managers, customer‑support staff and sales‑focused tasks also see solid agent performance. Results drop sharply for work that depends on deep, insider‑style domain knowledge.
 

A key caveat the lecturer emphasizes: strong results only show up when all implicit background knowledge held by a human gets written explicitly into prompts. Leave out that contextual information, and performance drops noticeably. A great deal of real‑world professional work involves figuring out which problems need solving in the first place, not just executing well‑laid‑out instructions. That critical piece of work never gets captured within GDP‑VAL.
 

This points toward the likely near‑term workflow: human‑agent collaboration. Humans handle high‑level problem framing, define objectives and supply missing context. Agents execute well‑defined subtasks under human oversight. This combination already delivers cost savings for many professional workflows, though suitability shifts a lot from one occupation to the next.
 

One student asks how seriously we ought to treat the dollar‑value labels attached to benchmark tasks. The lecturer replies we should not fixate too much on raw dollar figures. The real value of GDP‑VAL lies in building a structured task taxonomy tied to real‑world occupations. Before this benchmark existed, conversations about AI’s workplace impact stayed mostly qualitative and anecdotal.
 

DeepScholarBench: Evaluating Deep Research‑Synthesis Agents

 

The final benchmark covered is Stanford‑built DeepScholarBench. It targets deep research‑synthesis workflows, with its core assignment being drafting the related‑work section of academic papers. This problem resonates with many students in attendance, since writing literature reviews is a familiar, time‑consuming research pain point.
 

The benchmark draws recent arXiv papers across 22 PhD‑level research domains. It refreshes every month with newly published papers. To avoid training‑data contamination, it only uses papers released after major model training cutoff dates. Evaluation happens along three axes: knowledge‑synthesis quality, retrieval quality, and citation verifiability.
 

Knowledge‑synthesis scoring judges whether written output reads coherently, follows logical organisation, and captures key facts from source literature. Retrieval quality assesses how relevant and important the retrieved papers turn out to be. Verifiability checks whether each cited source genuinely backs up the specific claim the agent attaches to it. Human annotators reached 70‑80 % inter‑rater agreement for these evaluation criteria.
 

The most striking finding: none of today’s research‑focused agents exceed an overall 19 % score. This benchmark leaves massive room for improvement, and the lecturer notes it would make a solid starting point for student final projects. Performance splits unevenly across evaluation axes. Systems such as OpenAI Deep‑Research produce smooth, well‑organised writing yet regularly overlook critical facts. Retrieval quality stays mediocre overall. Agents struggle to pull in the most foundational papers within a given research field. Even when you hand an agent the perfect list of source papers up front, it still only captures roughly half of all key facts. A persistent trade‑off emerges: systems that generate polished prose often suffer poor citation verifiability, whereas systems prioritizing factual citations produce less elegant synthesis text.
 

Common failure modes include incomplete source discovery, missing foundational papers, weak ability to judge paper importance, incomplete extraction of key facts, and that ongoing tension between synthesis polish and verifiability. Expert human researchers carry years of accumulated domain knowledge in their heads, something agents cannot replicate. Agents have to locate and process every relevant document from scratch.
 

One student wonders why spend lecture time on such a narrow academic‑writing benchmark. The lecturer explains that research‑synthesis work serves as a representative proxy for a broader category of real‑world jobs. Financial analysis, investigative research and many analyst‑style assignments all demand reading large knowledge bases and assembling coherent output from scattered sources. DeepScholarBench gives us a concrete test bed for this whole class of problems.
 

Human researchers spend anywhere from thirty minutes up to eight hours on these synthesis assignments, placing them firmly into the long‑horizon agent‑evaluation bucket. Even if an agent hits METR’s 50 %‑success time‑horizon for this class of task, the actual output quality frequently remains sub‑par. Simply stepping through every task step does not equal producing trustworthy, high‑quality synthesis work.
 

Cross‑Benchmark Takeaways & Open Unresolved Questions

 

Each benchmark carries its own built‑in limitations. METR relies on automated scoring, focuses on single‑agent workflows, does not heavily penalise mistakes and ignores practical real‑world resource constraints. GDP‑VAL works with fully specified, one‑shot tasks. It cannot model iterative back‑and‑forth refinement, and it cannot measure work that hinges on unwritten, implicit domain knowledge. DeepScholarBench highlights real‑world gaps around source retrieval, fact extraction and citation trustworthiness.
 

Across all three benchmarks, agents perform best on isolated, clearly‑defined tasks within domains well‑represented in training datasets. Software engineering and machine‑learning‑adjacent research show the clearest progress. Even when agents produce long, well‑structured output, sections of that output can still contain errors.
 

Significant weak points persist. Performance drops sharply whenever critical context is missing from prompts. Agents struggle with ambiguous instructions that require them to work out what problem actually needs solving. Consistently hitting 95 % success reliability remains out of reach. Generalisation outside digital knowledge‑heavy work stays limited. Finding comprehensive high‑quality sources, extracting key facts, and keeping citations verifiable through multi‑step workflows all remain tough challenges.
 

Combining observations from all three benchmarks creates a nuanced picture. METR demonstrates exponential growth in maximum feasible task length at 50 % reliability. GDP‑VAL’s expert‑comparison results show slower linear improvement. DeepScholarBench illustrates that completing all procedural steps of a task does not guarantee high‑quality knowledge synthesis.
 

One part I found tricky to unpack is what these mixed trends mean for long‑term forecasts. METR‑based projections suggest agents might one day handle month‑long assignments. Results from the other two benchmarks give plenty of reason to tone down that optimism. Being able to finish a long‑running task does not automatically mean you get usable, professional‑grade output.
 

Multiple open questions surface during the Q&A segment. One student asks for clarification on METR’s 50 % success metric. It is not a requirement for every individual task to hit 50 % pass‑rate. Each separate task gets its own success rate measured across repeated agent runs. Researchers then aggregate those numbers across the full benchmark suite.
 

Another student asks what fundamental obstacles hold back performance on the long‑tail of hard tasks. The lecturer answers it stems from a mix of data‑related hurdles and fundamental model‑capacity constraints. Building realistic, representative task environments itself proves difficult. Some domains such as legal and financial research see steady iterative improvements. Fields like embodied robotics still devote huge effort toward data collection to close existing capability gaps.
 

Near the lecture’s end, the lecturer shares their personal perspective. At this moment in time, AI functions best as a collaborative brainstorming partner, rather than a fully independent AI scientist capable of completing full hypothesis‑experiment research loops. Achieving rock‑solid reliability across the long tail of hard assignments will turn out to be extremely challenging. Computer‑science‑focused professions are not going away anytime soon, partly because human workers still need to supply missing context and properly frame problems for AI agents.
 

The whole lecture circles back to agentic evaluations. We cannot draw conclusions from a single benchmark metric alone. We need multiple assessment lenses: task time‑horizon, comparison against real‑world expert work, plus synthesis quality and citation trustworthiness. Even so, all these metrics still need grounding against human evaluators.

Content Disclaimer: This article is for general reference only and does not constitute professional R&D guidance, production process advice or quality certification. All material performance data has specific test premises; readers should verify parameters against actual equipment and working conditions.