CS329A Self-Improving AI Agents, Part 8: Agentic Evaluations and Long-Horizon Tasks

CS329A’s eighth lecture asks what an agent evaluation should measure once a chatbot can produce fluent answers. It compares three perspectives: METR-style time horizons, GDPVal’s professional deliverables and DeepScholar-Bench’s research synthesis.

These perspectives ask different questions. How much human work does a task represent? Does an expert prefer the delivered artifact? Does a research report retrieve important sources and support its claims? No single number answers all three.

The lecture’s historical results are most useful as examples of measurement design and failure analysis. They are not a current ranking of models, and a favorable benchmark score should not be converted into a blanket claim of reliable autonomous work.

The time-horizon approach measures agent success across tasks whose difficulty is anchored by skilled humans’ completion times. A fitted relationship then describes the human-duration task associated with a chosen success level, such as 50% or 80%.

07:30View full image

That is not the same as the model’s elapsed running time. A task that takes a professional an hour may take an agent a different amount of time, and the horizon summarizes a set of observations rather than guaranteeing every task below that duration will succeed.

This definition also makes reliability part of the statement. A 50% horizon describes a much weaker standard than near-certain completion. Omitting the percentage can make the same graph sound more operationally reassuring than it is.

The lecture discusses collecting completion times from experienced professionals and accounting for variation among them. Even a careful baseline depends on who knows the environment, how the task is specified and what counts as a successful attempt.

Repository maintenance illustrates the issue. Someone embedded in a codebase may solve a problem much faster than an outside contractor who first needs context. Comparing an agent with either person answers a different practical question.

Messy tasks and unfamiliar repositories therefore deserve attention beyond a single fitted trend. A historical doubling pattern can describe progress in the evaluated suite without becoming a law about all work or a precise forecast of future month-long autonomy.

The reliability comparison shows a substantial gap between lower and higher success thresholds. The lecture’s exact examples belong to the models and studies being discussed; their purpose is to show why a capability horizon cannot stand alone.

Failure categories include poor planning, unsuitable tool choices, reasoning mistakes, abandoning a task too early and repeating an action after it has already failed. These are different defects and suggest different interventions.

A final answer that sounds complete is not enough to diagnose them. Evaluators need to inspect whether the requested work was completed, where the trajectory diverged and whether the system recovered. That connects benchmark measurement to the practical meaning of finishing a task.

GDPVal asks whether an output is good enough relative to work produced by experienced professionals. The lecture describes digital tasks across occupations, with artifacts such as documents, spreadsheets, designs or other files and often supporting reference material.

This broadens evaluation beyond a short question-answer pair. Correctness can include following instructions, using supplied data and producing a usable deliverable in the requested form. Human pairwise preference is useful where several valid outputs may exist.

A win on such a task does not establish that an entire profession can be replaced. Selecting priorities, acquiring tacit context and deciding which problem matters may not be included in the supplied prompt. Capability on the artifact is also separate from whether a real deployment is appropriate or authorized.

GDPVal’s win-rate chart and a time-horizon chart have different axes and outcome definitions. A student explicitly raises this point, and the lecturer agrees that the metrics should not simply be treated as disagreeing.

Wins and ties also need to be distinguished, while aggregate averages can hide large differences between task types. A model that formats a presentation well may make calculation or instruction errors elsewhere; one average cannot describe every deliverable.

The comparison is therefore complementary. Duration asks about a task scale at a reliability level; professional preference asks about output quality under its evaluation conditions. Neither licenses replacing careful task-level analysis with an appealing shape of a trend line.

The lecture highlights a failure in which a model promises to consult reference data but does not actually use it. A polished artifact can then contain invented information despite a confident account of the process.

Other failures concern formatting or instruction following, and human graders can disagree about what is acceptable. Those details matter to the interpretation of both win rates and claims about cost or speed.

A well-specified prompt can help a model perform a chosen task while leaving the discovery of the right task to a human. The evaluation’s context is part of the result, not a free resource that every workplace deployment automatically receives.

DeepScholar-Bench examines a more specific task: produce a related-work section by finding references, extracting relevant information and synthesizing claims with citations. The lecture describes recent papers and periodic refreshes as ways to reduce staleness.

It separates knowledge synthesis, retrieval quality and verifiability. A report can be organized yet miss key facts; retrieved papers can be relevant yet omit foundational work; citations can look credible yet fail to support the attached statement.

Freshness helps with contamination risk but is not a proof that no leakage exists. Likewise, a citation count can be a proxy for importance without establishing that a paper is correct or that a particular claim is supported. Each dimension needs its own check.

Citation precision asks whether the citations used support their associated claims. Coverage asks whether the claims that need support actually receive it. A report can be strong on one and weak on the other.

55:00View full image

The lecture describes systems that write fluent synthesis while missing important information, and experiments in which providing good sources still leaves incomplete fact extraction. Retrieval alone therefore cannot account for every failure in a research report.

Human validation of the evaluation metrics is also not perfect agreement on every judgment. The reported scores are evidence under that procedure, not an absolute measure of all scholarship or a guarantee that a report with many citations is ready to publish.

A research task can be long enough to appear in a time-horizon discussion, economically useful enough for a professional-work benchmark and still weak in source coverage or citation fidelity. Those statements can all be true at once.

The lecture connects the gaps to context management, multi-step information gathering, recovery and verifiable synthesis. It does not reduce every long-tail failure to one missing dataset or one model feature.

A practical evaluation should state the task, the supplied context, the success criterion, reliability, cost and evidence quality. Then a system’s improvement can be tied to what actually changed. That is more informative than asking whether it crossed one universal threshold called intelligence.