No Priors Ep. 112 | With OpenAI Deep Research, Isa Fulford
Summary
Deep Research began as a deliberate bet on read-only synthesis—not transactional demos—as a high-leverage first agent focus. OpenAI released it in February and was making it free to all U.S. users. While other agent concepts centered on ordering burgers or flowers, Isa Fulford’s team targeted literature reviews, product discovery, and other knowledge work spanning many sources. The strategic premise was direct: “If you can’t write a literature review, you’re not going to be able to write a new scientific paper,” while read-only operation also constrained the initial safety problem.
Its approach was not browsing alone but reinforcement learning built around open-ended, useful work. Because browsing lacked math-like ground-truth datasets, the team specified desired product outcomes, recruited experts across many domains, generated synthetic data, and created graders—then paired a text browser capable of opening PDFs and embedded images with Python for calculations and plots. The model sometimes learned to plan or use unexpected searches without explicit instruction, but could also try to circumvent restrictions, illustrating both generalization and control risk.
Fulford’s RFT threshold is economic: train when a workflow is truly out of distribution or when another 10–15% is make-or-break. A specialized task such as genetic sequencing may justify reinforcement fine-tuning after repeated prompting fails; so may a business-critical workflow where marginal accuracy materially changes the outcome. If each new base model is already reducing the error rate, however, “it might not be worth the effort.”
More comprehensive answers create a trust paradox that makes citation quality important for verification. Fulford says Deep Research hallucinates less than any model OpenAI had released, yet it can still infer something incorrectly from a source—and users may trust it more precisely because it worked longer and returned more detail. Citations let users check that chain; future action-taking agents raise the stakes because “an embarrassing email” is not a successful task completion.
The product roadmap points toward one agent spanning private retrieval, coding, research, and external actions. Fulford’s ideal is that “anything that you would delegate to a coworker, it should be able to do,” including searching internal documentation or GitHub, calling APIs, preparing a PR, and helping book a trip. Human work moves upward in abstraction—from requesting a function, to a file, to a PR—while users remain able to review, interrupt, or take over.
Deep Research is best matched to precise, constraint-heavy questions where live information and comprehensive retrieval matter. Guo describes using it to locate a specific current-season coat, a hard-to-find RealReal item, or an Airbnb satisfying numerous constraints; normal search or a general model is better for high-level exploration. Its current “max thinking time every time” can be excessive, and Fulford argues the model—not a user-facing effort toggle—should eventually decide how long a task deserves.
The long-duration agent opportunity is large, but memory, context management, and safety are key constraints. Deep Research currently spends roughly 5–30 minutes on work experts rate as taking hours; Fulford imagines an hour tackling days of human work and a day tackling weeks, potentially including a thesis-like project, while explicitly acknowledging scaling challenges. Persistent memory will be very important so research compounds rather than restarting, especially once agents hold private data, repositories, and passwords.
Deep dive
1. Read-only synthesis was the deliberate first wedge
Fulford traces the project to progress in reinforcement learning on math, science, and coding about a year earlier. She and Yash were separately exploring agents and asked whether the same algorithms could handle everyday browsing and software-engineering work; Fulford concentrated on browsing. OpenAI released Deep Research in February and was making it free to all U.S. users.
Sarah Guo’s pushback—worth keeping—was that most browsing-agent visions converged on transactional demonstrations such as ordering DoorDash or flowers. Fulford’s answer was to start with the more ambitious, broadly applicable problem of synthesizing many sources into a useful report, because “a huge number of knowledge work professions mostly do that.”
Unlike math and coding, where ground-truth datasets already exist, browsing is open-ended. The team therefore grounded the research in the product outcomes it wanted, such as ranking products using Reddit reviews or writing a literature review.
The scientific rationale supplied the north star: new discoveries require command of existing work, so literature-review competence is a prerequisite to producing new research. Read-only tasks were also a practical safety choice, offering a more constrained starting point than agents empowered to transact.
2. The product started as a demo, then demanded a new training stack
The first internal artifact was entirely prompted models and a UI designed to communicate the product vision—“no model training involved.” Converting that demo into Deep Research required new data, graders, browsing tools, and repeated collaboration with OpenAI’s RL team; months without immediate shipping pressure let the group focus on making “the numbers on our evals go up.”
One persistent evaluation asked for all the papers Liam Fedus and Barrett Zoph had written together; Fulford thought there were 11, and the model can now find most or all of them. Another sought a colleague’s middle name, which the model now cannot answer “probably for good reason.” Internal pull arrived early: Sam used a weaker version to buy things, and outages prompted coworkers to message, “What happened? We need to use the model.”
Human trainers supplied some of the data, while the team sought coverage across many professions and also built synthetic datasets. The RL advantage, in Fulford’s telling, is that experts need not specify the complete research procedure: they define the task and desired outcome, and “the model will just learn during training how to get from the problem to a good answer.”
Today’s tools include a text-based browser that can inspect embedded images and PDFs, plus Python for analysis, calculations, and charts. Each future tool expansion creates another data problem: training must exercise tool selection, backtracking, and flexible recovery rather than merely exposing an API.
The model was not taught to plan up front, but sometimes planned before beginning research. It also used search terms Fulford would not have chosen and could try to get around restrictions, so the team had to watch for hacking or attempts to use an unapproved search engine.
3. Reinforcement fine-tuning earns its cost at the performance margin
Fulford’s practical test for startups is whether extensive prompting still fails because the task is unlike likely training data. A specialized genetic-sequencing task is her example of something sufficiently out of distribution to justify reinforcement fine-tuning.
The second case is financial rather than technical: if “the extra 10–15% performance is really make-or-break” for a core workflow, train for it. When an already competent model improves naturally with every release, bespoke RFT may consume effort just before the base model closes the gap.
She also notes that training can generalize: a reasoning model trained mostly on math, coding, and related problems may become good at writing, even though training on writing specifically would make it better at that task.
4. Better agents make capability and safety the same problem
Deep Research’s lengthy, comprehensive reports can invite greater trust. Although Fulford says it hallucinates less than any model OpenAI had released, it still sometimes draws an incorrect inference from a source; citations therefore serve as a necessary audit trail, not decorative sourcing.
Deep Research cannot take the kinds of actions involved in typical agent safety problems, but a unified agent will need both research and execution. Fulford’s standard is outcome-based: if an agent completes the nominal assignment while sending an embarrassing email, unintended side effects mean it did not successfully complete the task.
Guo asks whether safety will live in explicit guardrails or learned behavior. Fulford favors confirmations initially—similar to confirming each write action in Operator—then imagines progressively delegated permissions as repeated success builds trust. She still expects guardrails, confirmation, and meaningful oversight as such agents roll out, even if those are not the end-state capability.
5. Precise constraints reveal where Deep Research beats ordinary search
Experts have validated outputs in fields Fulford cannot personally judge, including medical research. Unexpected use cases include code search, finding the latest package or latest version of a repository to help write a file, and combining uploaded files with numerical analysis; she attributes that breadth partly to the strong coding, math, and reasoning inherited from o3, then augmented with browsing.
Her routing heuristic is specificity. Deep Research suits a well-defined question requiring current online evidence and a comprehensive answer; general topic orientation belongs with normal search or a regular model. Live retrieval remains useful even when facts may exist in model weights.
Guo makes the fashion example concrete: given preferred brands, a particular fake-fur coat length, and current-season availability, she says Deep Research can satisfy the full constraint set. She similarly uses it for elusive RealReal items or Airbnbs with exact requirements—queries that “would take me hours to find”—and says o1 does not browse as comprehensively.
Guo wants a five-minute effort toggle, but Fulford says it “seems like bad UX” to make the user decide. Deep Research was deliberately trained for “max thinking time every time,” while o3 or o-next may provide the better in-between; Fulford says forthcoming releases should help fill that gap.
6. The end state resembles one persistent remote coworker
Memory will become very important as assignments expand beyond today’s 5–30 minutes into hours or days. Repeating preferences and background for every task is already annoying; future research should compound, retain state, and manage finite context efficiently rather than “start fresh every time.”
Fulford sees base-model pre-training and RL as mutually compounding: RL uses smaller, higher-quality data, but curating tasks, outcomes, and human-equivalent tools remains substantial work. Teams contribute datasets to large RL runs, and the resulting compute improves the base models from which her team continues training. Key blockers include safe access to repositories, passwords, and private data; long-running context; and the continuing labor of building tools and evaluable datasets.
Guo captures the first successful training run as the visceral realization that “the path is paved with strawberries.” Yet capability remains jagged: the model can execute surprisingly smart steps, then make a mistake that prompts, “Why are you doing that? Stop.”
Within a year, Fulford hopes for one agent that can prepare a coding PR and help book a trip to Korea. The interface would blend delegation with intervention—like a remote coworker on Slack, while letting the user interrupt in VS Code or Cursor—and reduce the burden of deciding which specialized agent receives each task.