Day 24 — Bonus 7 May 5, 2026

Deep Research Agents

How an LLM stops being a chatbot and becomes an analyst — the orchestrator + subagent pattern that now powers ChatGPT Deep Research, Anthropic Research, Perplexity, and Gemini.

Read: ~22 min Level: Senior engineer accuracy Day 24 of 24 (Bonus 7)
Curriculum progress
24 / 24
01 — Core Concept

What "deep research" actually is — and isn't

Throughout 2024, "agent that browses the web" mostly meant a single LLM that could call a search() tool. That works for "what's the weather in Tokyo." It does not work for "compare the GTM strategies of the top 5 European AI compliance startups, with citations." For the second kind of question, by early 2025 a new architecture had crystallized — and every major lab shipped a product version of it within four months of each other. OpenAI's Deep Research launched February 2, 2025. Perplexity Deep Research followed February 14. Google's Gemini Deep Research had been in preview since December 2024. Anthropic's Research mode shipped in April 2025 and the team published a remarkably candid engineering write-up in June. The pattern they all converged on is what this lesson is about.

The analogy first. A chat agent is a junior associate you Slack-message — fast, conversational, single-shot. A deep research agent is a research director who delegates to a small team. The director reads your prompt, decomposes it into sub-questions, hires three or four interns to chase down each thread in parallel, gets their findings back, decides whether more digging is needed, and then writes the final report with citations. The director never reads the raw web pages — that's the interns' job. The director reads summaries. This split is the whole game.

Now precise. A deep research agent is a multi-stage system in which a lead agent (also called orchestrator, coordinator, or planner) decomposes a user query into research subtasks, dispatches them to subagents running in parallel, each of which executes its own ReAct-style loop over web tools (search, fetch, read), returns a structured finding to the lead, and the lead either iterates (more subagents) or synthesizes (writes the report). The August 2025 survey by Zhang et al. (arXiv 2508.12752) formalizes this as a four-stage pipeline: planning → question developing → web exploration → report generation. Anthropic's June 2025 engineering blog reports their multi-agent system with a Claude Opus 4 lead and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on their internal research eval — the most-cited number in the field.

1Three properties that distinguish deep research from RAG

Open-ended retrieval. RAG (Day 4) retrieves from a known, indexed corpus. Deep research retrieves from the open web — the search space is unbounded, sources are heterogeneous (HTML, PDF, JSON, paywalled), and freshness matters. The subagent must decide what to search for next based on what it just read.

Multi-turn iterative refinement. A deep research run averages 30-50 search calls and 100-200 page reads, executed across 5-30 minutes of wall-clock time. Per Anthropic's blog, their research agents use roughly 15× more tokens than a normal Claude chat. This is not a performance bug — it's the product. You're paying for breadth and depth.

Citation-grounded synthesis. The output is a structured report with inline citations to specific URLs. Hallucinated citations (a recurring failure mode) are a P0 bug, not a quality issue. This forces a discipline: the synthesizer has to be able to say "I don't have evidence for X" rather than fill in plausible-sounding text.

Mental model: RAG = "look up the answer in the company wiki." Deep research = "open Chrome, search, click, read 80 tabs, and write me a memo."

2Why parallel subagents beat a single long-context call

The naive alternative — stuff every search result into a 1M-token context window and let the model reason over it — sounds appealing and mostly does not work. Two reasons.

Attention dilution. Performance on retrieval-from-long-context degrades non-linearly with context length. The "needle in a haystack" plots look great; the "two needles in a haystack" plots fall off a cliff. Real research questions need 5-20 distinct findings, not 1.

Compute parallelism. Five subagents browsing in parallel finish in roughly 1/5 the wall-clock time of one subagent doing all the work serially. For products charging for 5-30 minutes of agent time, this is the difference between shipping and not shipping.

The Anthropic blog quantifies this — their internal eval credits "token-budget-allocation across multiple agents" with the bulk of the 90.2% improvement, not better prompting of any single agent.

The four-stage pipeline — in one diagram

┌──────────────────────────────────────────┐ │ STAGE 1 · PLAN │ // lead agent │ user query → research plan + sub-tasks │ └──────────────┬───────────────────────────┘ │ ▼ ┌──────────────────────────────────────────┐ │ STAGE 2 · QUESTION DEVELOPING │ // lead agent │ decompose → 3-8 specific sub-queries │ └──────────────┬───────────────────────────┘ │ ▼ spawn N subagents in parallel ┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐ │ SubA-1 │ │ SubA-2 │ │ SubA-3 │ │ SubA-N │ // each runs ReAct loop │ search │ │ search │ │ search │ │ search │ // search(q) │ fetch │ │ fetch │ │ fetch │ │ fetch │ // fetch(url) │ read │ │ read │ │ read │ │ read │ // read(content) │ summz │ │ summz │ │ summz │ │ summz │ // summarize → return └───┬────┘ └───┬────┘ └───┬────┘ └───┬────┘ │ │ │ │ └───────────┴─────┬─────┴───────────┘ ▼ STAGE 3 · WEB EXPLORATION (returns) ┌──────────────────────────────────────────┐ │ STAGE 4 · REPORT GENERATION │ // lead agent (or writer) │ N findings → cited synthesis report │ │ (gap detected? loop back to Stage 2) │ └──────────────────────────────────────────┘

Two architectural details worth burning into memory. First, the lead agent never reads raw web pages — it reads structured findings (typically 200-500 tokens per subagent return). This keeps the lead's context window healthy across 20+ subagent invocations. Second, the loop in Stage 4 ("gap detected → re-plan") is what makes "deep" research deep. A single shot is not enough — the lead needs to look at the assembled findings and notice "I have nothing about pricing" or "two sources contradict each other on revenue," then spawn another wave of subagents.

~15×
Tokens per query vs. normal chat (Anthropic, June 2025)
90.2%
Multi-agent uplift over single-agent on Anthropic's internal eval
5-30 min
Wall-clock time per OpenAI Deep Research run

02 — Papers to Know

Four primary sources that define this field

SURVEY · 2025 arXiv preprint 2025
Deep Research: A Survey of Autonomous Research Agents
Wenlin Zhang, Xiaopeng Li, Yingyi Zhang, Pengyue Jia, Yichao Wang, Huifeng Guo, Yong Liu, Xiangyu Zhao · arXiv:2508.12752 (Aug 2025)
The first systematic taxonomy of the deep research pipeline. Defines the four canonical stages — planning, question developing, web exploration, report generation — and surveys 80+ implementations against this skeleton. For each stage, it catalogs the technical challenges (e.g., decomposition granularity, search-query reformulation, source ranking, citation accuracy) and the methods proposed to address them. If you read one paper, this is the one — it's the rosetta stone for the rest of the literature.
Why it matters: Before this survey, every product team described their deep research pipeline differently. The four-stage taxonomy now anchors how teams compare implementations, decompose failures, and prioritize improvements. "Stage 2 failure" is now a thing engineers can say in standup.
arxiv.org/abs/2508.12752
SEMINAL · 2025 arXiv preprint 2025
BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents
Jason Wei et al. (OpenAI) · arXiv:2504.12516 (April 2025)
1,266 questions that require persistent multi-hop web navigation to answer — questions like "What was the first paper to cite this 2003 conference proceedings on graph algorithms?" Answers are short and verifiable; the difficulty is in the navigation. GPT-4o without browsing scores 0.6%; with browsing it climbs only to 1.9%. The early Deep Research model was the first system to break double digits. The benchmark now anchors the entire field — every paper in this list reports a BrowseComp number.
Why it matters: Made deep research evaluable. Before BrowseComp, "is my deep research agent good?" was a vibe check. After BrowseComp, it's a number you can put on a dashboard and a value you can train against with RL.
arxiv.org/abs/2504.12516
RECENT · 2025 arXiv preprint 2025
Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning
Bowen Jin et al. · arXiv:2503.09516 (March 2025)
Built on the veRL framework, Search-R1 trains a model to interleave reasoning steps with search-engine calls, with rewards shaped from final-answer correctness. On seven QA datasets, it improves performance by 41% on Qwen2.5-7B and 20% on Qwen2.5-3B over RAG baselines. The takeaway is that you do not have to prompt-engineer a search agent — you can train one, the same way DeepSeek trained R1, just with the search tool inside the loop.
Why it matters: Showed that "agentic search" is a fine-tunable skill, not a prompting trick. Opened the door to the wave of RL-trained search agents that followed (R1-Searcher, WebDancer, WebSailor) — the recipe that now powers most open-weight deep research models.
arxiv.org/abs/2503.09516
RECENT · 2025 arXiv preprint 2025
WebDancer: Towards Autonomous Information Seeking Agency
Tongyi Lab, Alibaba · arXiv:2505.22648 (May 2025)
A four-stage post-training recipe that takes an open-weight base model and turns it into an autonomous web-research agent: (1) browsing-data construction, (2) trajectory sampling, (3) supervised fine-tune cold start, (4) RL on the agent loop. Evaluated on GAIA and WebWalkerQA, the result surpasses GPT-4o in best-case scenarios — using a much smaller open-weight backbone. Spiritually a sibling to Search-R1, but at full agent-trajectory granularity rather than just search-call granularity.
Why it matters: Together with WebSailor (its successor), this is the playbook for building a state-of-the-art deep research agent without OpenAI's data or compute budget — and the work behind why Tongyi Deep Research is now genuinely competitive with the closed frontier on browsing benchmarks.
arxiv.org/abs/2505.22648
ENGINEERING BLOG · 2025
How we built our multi-agent research system
Anthropic Engineering · June 2025
Not a paper — but the most-cited engineering write-up in the field. Documents the lead-researcher-plus-subagents architecture behind Claude's Research mode, the prompt-engineering lessons (vague subagent objectives → duplicated work, e.g., one subagent investigating the 2021 chip crisis while two others duplicate effort on 2025 supply chains), and the headline 90.2% multi-agent uplift over single-agent on their internal eval. The clearest free public explanation of how a frontier-lab deep research agent actually works under the hood.
Why it matters: Engineers reading this blog will recognize 80% of the design choices in their own products. Anthropic published it specifically because they noticed teams kept reinventing the same wheel poorly.
anthropic.com/engineering/multi-agent-research-system

03 — GitHub Pulse

The open-source deep-research stack as of May 2026

Within sixty days of OpenAI launching Deep Research, every major OSS framework had its own implementation. Within six months, Tongyi Lab had open-sourced a full training pipeline that closes most of the gap to closed-source. The repos below are where to look first.

huggingface/smolagents ~26K ★
A barebones library for agents that "think in code"; ships an open_deep_research example that hit 55% pass@1 on GAIA in a 24-hour replication of OpenAI Deep Research.
Architectural interest: code-as-action — the agent emits Python that orchestrates tools, instead of JSON tool calls.
dzhng/deep-research ~19K ★
The simplest possible deep-research agent — <500 lines of TypeScript with adjustable breadth and depth knobs. The reference "what's the smallest thing that works" implementation.
Architectural interest: shows that the recursive search-deepen loop is doable in tens of lines, before any framework abstraction.
langchain-ai/open_deep_research ~2.7K ★
LangGraph-based, configurable, multi-provider. A planner constructs a section-by-section research plan; each section is researched in parallel and written by a writer agent.
Architectural interest: section-as-subagent decomposition — easy to map to long-form output structure.
Alibaba-NLP/DeepResearch ~5K ★
Tongyi Lab's full open-source stack — WebWalker (benchmark), WebDancer (RL agent), WebSailor (post-training recipe), WebShaper (data synthesis). The most complete training pipeline that exists outside a frontier lab.
Architectural interest: end-to-end recipe — data construction, trajectory sampling, SFT cold start, agentic RL — that you can actually run.
PeterGriffinJin/Search-R1 ~3.6K ★
RL training framework for search-and-reason interleaved LLMs, built on veRL. The reference codebase for the Search-R1 paper.
Architectural interest: cleanest demonstration that you can RL a search loop — separate reward shaping for format, retrieval, and final answer.
langchain-ai/local-deep-researcher actively maintained
A fully local deep-research agent — runs against Ollama with a local search backend. The "no API keys, no data leaves the laptop" variant.
Architectural interest: shows what changes when you can't afford GPT-4 calls per subagent — heavier reliance on smaller local models with tighter prompts.

04 — Community Pulse

What practitioners are saying — last 60 days

Anthropic Engineering — June 2025 blog
"Each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries... vague instructions like 'research the semiconductor shortage' often were vague enough that subagents misinterpreted the task or performed the exact same searches as other agents — with one subagent exploring the 2021 automotive chip crisis while 2 others duplicated work investigating current 2025 supply chains."
The most-cited paragraph in the most-cited engineering blog of 2025. Source: anthropic.com/engineering/multi-agent-research-system
Harrison Chase (LangChain CEO) — 2026
Chase has argued that what makes "deep agents" deep is the sophistication of the looping architecture — planning, context management, memory, subagents, and richer prompting — and that the "harness" surrounding the model is now where most production value gets built, not the model itself. LangChain's bet on Deep Agents as a category (with Deep Agents UI, deep_research_from_scratch, and an Interrupt 2026 keynote) operationalizes this view.
Paraphrased from LangChain's 2026 blog series and Sequoia podcast appearance.
Simon Willison — Feb 2025 → Apr 2025
Willison's posts on Perplexity Deep Research (Feb 16, 2025) and "AI assisted search-based research actually works now" (April 2025) marked the inflection point in practitioner sentiment — the public moment when a generally-skeptical voice said the search-based research category had crossed from "almost works" to "useful for real work."
Source: simonwillison.net/2025/Feb/16/introducing-perplexity-deep-research/ and the April Substack follow-up.
Andrej Karpathy — 2025 LLM Year in Review
Karpathy framed 2025 as the year RLVR (Reinforcement Learning from Verifiable Rewards) became the de facto post-training stage, and noted Claude Code as "the first convincing demonstration of what an LLM Agent looks like — something that in a loopy way strings together tool use and reasoning for extended problem solving." Deep research agents are the search-loop sibling of this same recipe: the same RLVR machinery, applied to a verifiable-answer search environment instead of a code environment.
Source: karpathy.bearblog.dev/year-in-review-2025/

05 — Platform Deep-Dive

How the four big platforms implement this

Claude Research
Anthropic · Claude Opus 4.6 + Sonnet 4.6
Architecture: Lead-researcher (Opus) plus parallel subagents (Sonnet). Each subagent uses interleaved thinking and returns structured findings to the lead, which decides whether to spawn another wave. The 90.2% uplift number comes from this exact configuration.
Recent (last 30 days): Research mode is now exposed as a first-class capability in the Claude API for Cowork and developer integrations, with the lead/subagent split managed by the platform rather than the application.
OpenClaw Research Skill
Anthropic Cowork — skills + plugins
Architecture: Deep research is delivered as a skill plus a set of MCP tools (web search, fetch, computer use). The orchestration loop is in the skill's SKILL.md; subagents are spawned via Task / Agent calls with isolated context windows. This is the hub-and-spoke pattern from Day 14, applied to research.
Recent (last 30 days): tighter integration with the user's connected MCP servers (Slack, Drive, GitHub) — the research subagents can now pull from private corpora as part of the same plan.
ChatGPT Deep Research
OpenAI · o3-deep-research / GPT-5.4
Architecture: A version of o3 specifically optimized for browsing and analysis, plus a second o3-mini model used to summarize the chain-of-thought. Single-agent loop (not lead/subagent), but with a very long browsing trajectory — autonomous for 5-30 minutes per query, scoring 26.6% on Humanity's Last Exam at launch.
Recent: Now exposed via the o3-deep-research API model, lifting deep research from a ChatGPT-only feature to a programmable building block.
Gemini Deep Research
Google · Gemini 3.1
Architecture: Plan-first (the user reviews and edits the multi-step research plan before execution), then a long-running browsing trajectory inside Gemini's 2M-token context. Heavier reliance on long context than on parallel subagents — a different architectural bet from Anthropic's design.
Recent: Tighter integration into Workspace — research output flows directly into Docs and Sheets with source links preserved as cell-level citations.
Reading these four side-by-side: Anthropic bets on parallel subagents, OpenAI bets on a single very-long-trajectory agent, Google bets on long context plus user-in-the-loop planning, and OpenClaw delivers deep research as composable skills. Same problem, four genuinely different architectural answers — the best evidence we have that the "right" design is not yet settled.

06 — Vocabulary

10 terms to use precisely in conversation

Lead agent / orchestrator / coordinator
The top-level agent that owns the research plan, dispatches subagents, and writes the final report. Sees only structured findings, not raw web pages.
Trap: using "lead agent" interchangeably with "the LLM" — these are the same model weights but very different prompts, contexts, and roles.
"Our lead agent runs Opus; subagents run Sonnet — that mix is half the cost win."
Subagent / research worker
A short-lived agent that executes one specific sub-question with its own ReAct loop and tools, then returns a 200-500 token structured finding to the lead.
Trap: confusing "subagent" with "tool call." A subagent has its own context, prompt, and multi-turn loop; a tool call is a single function invocation.
"That subagent kept timing out — its sub-task was too broad."
Research budget / token budget
The total tokens / wall-clock time / search calls a single research run is allowed to consume. Allocated by the lead and divided among subagents.
Trap: confusing "research budget" with "context window." The budget is across many windows, summed.
"We hard-cap each subagent at 30K tokens; the lead gets 100K for synthesis."
BrowseComp
OpenAI's 1,266-question benchmark for multi-hop browsing agents (arXiv 2504.12516). The de facto evaluation for how well an agent can navigate the open web to find entangled facts.
Trap: treating BrowseComp as a research-quality benchmark. It measures information-seeking, not synthesis. A model can ace BrowseComp and still write a bad report.
"WebSailor-72B hits 12.0% on BrowseComp-en — strong for an open-weight system."
GAIA
A 466-question benchmark for general AI assistants spanning web, math, multimodal, and coding tasks. Older than BrowseComp and broader; remains the headline number for "is this agent good in general."
Trap: comparing two agents on GAIA validation vs. test split — they're not the same.
"smolagents' open_deep_research hits 55% pass@1 on GAIA validation."
Citation grounding
The discipline of attaching every claim in the synthesized report to a specific source URL, ideally with the supporting span quoted. The single biggest quality lever in deep research.
Trap: assuming the model "knows" not to hallucinate citations. Citation accuracy is a property of the system, not the model — needs explicit verification passes.
"Their report has 47 citations and 6 of them don't actually support the claim — that's a citation-grounding failure, not a writing failure."
Iterative refinement / research loop
The lead agent's "do I have enough?" decision after each wave of subagents, and the spawning of follow-up subagents to fill identified gaps.
Trap: hard-coding a fixed number of iterations. The point of iterative refinement is that the model decides when to stop.
"We saw a 3× quality jump when we let the lead spawn a second wave instead of capping at one."
RLVR for search
Reinforcement Learning from Verifiable Rewards, applied to a search-and-reason loop. Rewards based on final-answer correctness shape both the reasoning and the search-query generation. Search-R1 is the canonical example.
Trap: confusing RLVR with RLHF. RLHF rewards human preference; RLVR rewards verifiable correctness — much cleaner signal, but only available where ground truth exists.
"Once they switched from SFT to RLVR on the agent loop, retrieval relevance jumped 20 points."
Trajectory / rollout
The full sequence of (state, action, observation) tuples that one agent run produces. The unit of training data for agentic RL.
Trap: thinking of a trajectory as a "log." Logs are for humans; trajectories are structured training data with reward labels.
"WebDancer's trajectory-sampling stage generated 50K rollouts before the SFT cold start."
Subtask decomposition
The lead agent's transformation of a research question into 3-8 specific, non-overlapping sub-questions. The single highest-leverage prompt-engineering target in the whole pipeline.
Trap: too few sub-tasks (subagents do redundant searches) or too many (lead can't synthesize). The sweet spot is workload-dependent.
"The lead's decomposition was 'research the semiconductor shortage' — that's not decomposition, that's the original question."

07 — Expert Questions

Five questions that signal you've thought about this

Q1
When a subagent's findings contradict another subagent's, what's the lead agent's policy — re-spawn, prefer the more recent source, or surface the disagreement in the report?
Tests whether the candidate has actually shipped a deep research product. Naive systems silently drop one of the two; good systems either resolve via fresh evidence or explicitly preserve the disagreement. There's no universal right answer — but there is a wrong answer (silently picking one).
Q2
How do you bound subagent token budgets without starving the ones that need the most exploration? Static caps, lead-allocated quotas, or token markets?
A practical operations question with no settled answer. Anthropic mentions adaptive allocation; most OSS implementations use static caps. The candidate's answer reveals whether they've operated this in production or just read about it.
Q3
When does parallel sub-research lose to a single long-context call with all the web pages stuffed into 2M tokens? Where's the crossover for current models?
Probes the actual reasoning behind the architecture. Long-context performance plateaus on multi-needle retrieval; subagents win when the answer requires combining 5+ distinct findings. Below that, long context is faster, cheaper, and simpler.
Q4
What does "citation grounding" actually mean operationally? You've written the report — how do you verify each citation actually supports the sentence next to it?
Non-trivial. Real implementations run a separate verifier pass over (sentence, cited URL) pairs. Without that, you ship hallucinated citations. The candidate should reach for "verifier pass" or "NLI-style entailment check" — not "we trust the model."
Q5
If you had to pick one of (a) better models, (b) better tool design, (c) better orchestration prompts, (d) RL on the agent loop — which gives the largest deep-research quality lift in 2026, and why?
The Anthropic blog and the WebDancer/WebSailor results triangulate to the same answer: orchestration prompts + RL on the loop deliver more than swapping the base model. A candidate who answers "(a) better models" is anchored on 2024 mental models; the right answer is "(c) and (d) together."

08 — CGO Lens

What today's topic means for botlearn.ai

1Deep research is the highest-margin agent product right now

OpenAI charges $200/month for Pro (250 deep research queries) — about $0.80 per query. Anthropic's Claude Pro tier exposes Research at a similar implicit per-query margin. These products use ~15× more tokens than chat but customers gladly pay 5-10× more for them, because the output looks like analyst-grade work. For botlearn.ai's curriculum-generation use case, this margin profile is the most attractive in the agent product landscape.

2The buy-vs-build decision changed in Q1 2025

Through 2024, a deep research agent was a science project — no team without a frontier-lab background could ship one. By Q1 2025, OpenAI's launch + the smolagents 24-hour replication showed it was a 2-3 engineer-month project, given the right base model. By mid-2025, Tongyi's WebSailor open-sourced the full training recipe. For an education company that wants its own research agent (own brand, own data, own pricing), build is now realistic. The strategic question is whether your differentiation lives in the agent or in the corpus the agent reads.

3Cost has the same shape as a SaaS storage curve

OpenAI's o3-deep-research API was the first move that turned deep research from a consumer feature into a developer building block — meaning every B2B SaaS will start adding "research" buttons in 2026. The competitive question is not "do we have a research feature?" — within twelve months, every product in your category will. The question is what makes a research output uniquely defensible in your vertical. For botlearn.ai, two candidates: (a) curated source list specific to AI/agent education that no general-purpose research agent ranks, and (b) a research-output format that compiles into a didactic asset (lesson, deck, video script) — not a generic report.

4The China factor — directly relevant for an Asia-facing CGO

Tongyi Lab's WebSailor and the Alibaba-NLP/DeepResearch repo mean that for a China- or APAC-targeted product, a fully open-weight, self-hosted deep research agent is now a real option. WebSailor-72B reports 12.0% on BrowseComp-en and 30.1% on BrowseComp-zh — the Chinese-language number is competitive with the closed frontier. For botlearn.ai's Xiaohongshu / Chinese-market presence, an open-weight self-hosted research agent removes both the API tax and the data residency concern in one move.


09 — Curriculum Tracker

24 days · core 17 + 7 bonus deep-dives

D01 Full Agent Stack
D02 Memory Architecture
D03 Planning & Tool Use
D04 RAG Deep Dive
D05 Agent Frameworks
D06 Benchmarks & Eval
D07 Multi-Agent Systems
D08 Computer Use Agents
D09 Code Agents
D10 Long-Horizon Tasks
D11 Agent Safety
D12 Agent Economics
D13 Research Frontiers
D14 OpenClaw Deep-Dive
D15 A2A Protocols
D16 Agentic Commerce
D17 Synthesis
D18 Reasoning & Test-Time Compute
D19 Agent Observability
D20 Voice & Realtime Agents
D21 Agent RL & Fine-tuning
D22 Context Engineering
D23 Open-Weight Agent Models
D24 Deep Research Agents