Skip to content
Agentic Engineering Fieldbook

Research and Knowledge Work Agents

Research and knowledge work agents promise to search, read, compare, summarize, and synthesize faster than a person can do by hand. They are useful because much knowledge work is not a single retrieval problem. The system may need to reformulate queries, open sources, follow citations, compare claims, detect conflicts, and produce a memo that a person can act on.

That same flexibility creates the core risk. A research agent can produce a polished answer whose citations do not support the claims, whose sources are stale, whose uncertainty is buried, or whose missing evidence is invisible. In research work, fluency is often the least important part of the deliverable.

Thesis

Section titled “Thesis”

A research agent is useful when it is designed as an evidence triage and synthesis system: every material claim should be connected to inspectable sources, dates, scope, uncertainty, and gaps that a reviewer can challenge.

The accepted outcome is not “an answer.” It is a bounded answer with an evidence trail. The evidence trail should be compact enough to read and concrete enough to audit.

Start with the research decision

Section titled “Start with the research decision”

Research-agent task fit depends on the decision the work supports:

  • background scan for a team planning session;
  • literature search before deeper expert review;
  • market or competitor summary with dated source links;
  • policy or standards comparison;
  • due diligence memo that identifies sources and unresolved questions;
  • internal knowledge synthesis across documents.

These tasks differ in consequence. A low-stakes scan may accept broader uncertainty and faster turnaround. A legal, medical, financial, safety, or compliance question may allow the agent to gather and organize evidence, but the accountable conclusion must belong to a qualified person or process.

The product promise should name the level of work. “Find and summarize sources” is different from “recommend a regulatory strategy.” “Prepare a literature map” is different from “decide clinical relevance.” When the downstream decision is consequential, the agent should surface review needs rather than smooth them away.

Source discovery is a controlled loop

Section titled “Source discovery is a controlled loop”

A research agent usually needs an observe-decide-act loop:

  1. Interpret the question and identify volatile terms.
  2. Choose search strategies and source classes.
  3. Open sources and record access dates.
  4. Extract claims, dates, definitions, and limitations.
  5. Compare sources for agreement, conflict, and coverage gaps.
  6. Revise the search when evidence is weak or one-sided.
  7. Produce a synthesis with citations tied to specific claims.

OpenAI’s Deep Research materials and BrowseComp benchmark illustrate the need for persistent browsing and verifiable answers in difficult web tasks (OpenAI Deep Research System Card; BrowseComp). GAIA similarly tests assistants on multi-step questions that require reasoning and tool use beyond single-turn recall (Mialon et al., 2023).

The engineering lesson is not that a browsing agent is automatically reliable. It is that search and synthesis are stateful processes. The system must remember what it tried, what it found, what it rejected, and which claims remain unsupported.

Citations are claims, not decoration

Section titled “Citations are claims, not decoration”

A citation is useful only when it answers three questions:

  • Which claim does this source support?
  • What is the source’s date, version, and scope?
  • What limitation should the reader keep in mind?

Research-agent interfaces should not dump raw citations at the end and make the reader reverse-engineer the argument. The source should appear near the claim or in a traceable evidence table. A cited source should not merely contain similar words; it should support the proposition being made.

Source type matters. A product documentation page can establish what that product says it supports. A system card can describe a provider’s safety framing and known limitations. A peer-reviewed paper can support an empirical or theoretical claim under its conditions. A regulation or standard can establish an obligation or definition in scope. A blog post can preserve provider or practitioner experience. These are not interchangeable.

Handle uncertainty as a first-class result

Section titled “Handle uncertainty as a first-class result”

Research agents should have explicit states for:

  • insufficient evidence;
  • conflicting sources;
  • stale or undated sources;
  • inaccessible primary sources;
  • source concentration in one vendor or region;
  • unresolved terminology;
  • need for expert review.

NIST AI 600-1 emphasizes provenance, data lineage, validation, and lifecycle risk for generative AI systems (NIST, 2024). In research products, that translates into source trails, transformation notes, and uncertainty statements. The user should be able to see not only what the agent concluded, but also what the agent could not establish.

Uncertainty should not become boilerplate. “I may be wrong” is almost useless. Better statements are specific:

  • “I found vendor documentation for this API behavior but no independent implementation evidence.”
  • “The standard summary is public, but the normative text is paywalled.”
  • “The source was updated continuously and does not expose a stable publication date.”
  • “The strongest evidence comes from benchmark tasks, not production deployments.”

Design options

Section titled “Design options”
Option Best fit Main risk
Retrieval answer Narrow question over a trusted corpus Misses external changes or conflicting sources
Browsing research agent Open-web scan, current facts, source discovery Citation drift, source quality errors, stale pages
Corpus synthesis agent Internal documents, enterprise knowledge, policy comparison Permission leakage and unreviewed internal contradictions
Multi-agent research system Broad search space with parallel subtopics Coordination overhead and inconsistent citation standards
Expert-assist research workflow High-stakes review preparation Human reviewer may over-trust polished synthesis

Anthropic’s multi-agent research system article is useful because it shows the benefit and cost of splitting broad research across parallel agents (Anthropic, 2025). Parallelism can improve breadth, but it also increases the need for source normalization, contradiction handling, and final synthesis discipline.

Failure modes

Section titled “Failure modes”
  1. Citation laundering. The agent cites a credible source that does not support the specific claim.
  2. Search myopia. Early query terms anchor the search and exclude better source classes or opposing evidence.
  3. Temporal collapse. Old, future, draft, and current sources are mixed without dates.
  4. Authority confusion. Vendor documentation, law, research, standards, and community reports are treated as the same kind of evidence.
  5. Synthesis overreach. The agent turns limited evidence into a broad recommendation.
  6. Hidden gaps. Missing primary sources or inaccessible full texts disappear from the final answer.
  7. Unsafe personalization. The agent uses private or internal knowledge in a way the recipient is not authorized to see.

Fieldbook guidance

Section titled “Fieldbook guidance”

Design research agents around evidence obligations. Require a claim-source map for material claims. Record access dates for volatile sources. Preserve source type and limitation. Make the agent comfortable returning a partial result when the evidence is incomplete.

Separate research work from accountable judgment. The agent may gather, compare, and draft. A responsible actor still owns decisions that require professional expertise, risk acceptance, or organizational authority.

Keep the evidence trail useful. Raw browsing traces are too noisy for most readers, while citation-only answers are too thin. The best artifact is usually a synthesis with compact source notes, dated references, and explicit unresolved questions.

Checklist

Section titled “Checklist”
  • Does the product promise distinguish evidence gathering, synthesis, recommendation, and decision authority?
  • Are source classes, dates, jurisdictions, and versions recorded where they affect the claim?
  • Can each material claim be traced to a source that actually supports it?
  • Does the agent expose conflicting, stale, inaccessible, or one-sided evidence?
  • Are internal documents and private data scoped to the recipient’s authorization?
  • Does the workflow include expert review for high-stakes domains?
  • Are citations preserved near claims or in a usable evidence table?
  • Does evaluation test citation support, not only answer helpfulness?

References and how they are used

Section titled “References and how they are used”