Guide
Use AI agents for research by defining the question, source boundary, evidence packet, uncertainty, decision owner, and review loop.
AI agents can make research faster when they turn a bounded question and an approved source set into an evidence packet a teammate can inspect. The useful output is not a confident summary alone. It is a record of what the agent was asked to find, which sources it used, what those sources support, what remains uncertain or conflicting, and which person owns the next decision.
Commonly (commonly.me), the shared workspace where humans and AI agents work together, gives research work a durable collaboration surface: a pod can hold the question and its discussion; a task makes the owner and expected result visible; attachments carry a memo; shared memory can preserve an accepted conclusion for the next session. The workspace does not decide which source is authoritative or make a recommendation true. Teams still need a research scope, a claim boundary, and a reviewer who can judge the resulting evidence.
Research is a strong first agent role because the contribution can be narrow and reviewable. An agent may retrieve named material, extract relevant facts, compare sources, and prepare a brief. It should not silently convert incomplete evidence into a policy decision, publish a claim, or expand its access “just in case.” Those are separate decisions that belong to a human owner and the systems that enforce access.
This guide explains how to define an AI research role, design a source boundary, create an evidence packet, handle uncertainty, and evaluate the work before relying on it for a higher-consequence decision.
The point of research is to improve a decision. An AI agent can help collect, structure, and compare evidence. It does not inherit the authority to choose the team’s strategy, make an external commitment, or state a weak inference as fact merely because it found a plausible pattern.
| Research stage | Agent contribution | Human responsibility |
|---|---|---|
| Frame the question | Restate the assigned question and identify missing scope details | Decide which question matters and what decision it is intended to inform |
| Gather sources | Retrieve the named, authorized material and record what was consulted | Select authoritative source classes and approve any expansion of the source boundary |
| Extract evidence | Identify facts, relevant passages, conflicts, and gaps | Judge whether the evidence is sufficient for the consequence of the decision |
| Synthesize | Prepare options, patterns, or a clearly labeled recommendation | Choose the direction, accept tradeoffs, or request more research |
| Handoff | Attach an evidence packet and ask one bounded question | Accept, reject, narrow, or defer the next step |
This division is useful because it keeps the research result auditable. A person can ask, “Which source supports this claim?” or “What would change this recommendation?” An agent that cannot answer those questions has produced prose, not a dependable research artifact.
“Research AI agent security” is not a research task. It is a topic. A decision question gives the agent a way to identify relevant evidence and gives the reviewer a way to tell whether the result is complete.
| Broad topic | Better decision question |
|---|---|
| AI agent security | Which permission and review boundaries are required before a public-facing agent can access this workspace? |
| Customer feedback | What recurring issue appears in the approved support reports, what evidence supports it, and what information is still missing before triage? |
| Product documentation | Which claims in this draft are supported by the named documentation, and which need a source, revision, or removal? |
| Technical direction | What do the designated project records say about the proposed integration’s scope, dependencies, and unresolved decision? |
| Competitor landscape | Which explicitly named public sources support the comparison criteria, and where is the evidence too thin to make a claim? |
A strong research question usually has four elements:
If one of these is missing, the research agent should ask a focused clarification or record the gap. It should not select a decision owner, invent a time frame, or broaden an ambiguous topic into a large, unreviewable search.
The source boundary is the most important instruction in a research role. It says what the agent may treat as evidence, what it must not access, and how it should react when the available sources cannot answer the question.
| Source rule | Why it matters |
|---|---|
| Name the approved sources or source classes | A reviewer can reproduce the evidence path and identify a missing authority |
| Specify current-state sources | The agent knows to read the live task, focused thread, or accepted decision before acting on an old message |
| State whether external material is in scope | The agent does not silently widen the research surface |
| Mark private and sensitive material as out of scope | Access follows the role’s need, not the agent’s curiosity |
| Define how to handle conflicts | The agent presents competing evidence rather than choosing a winner without authority |
| Define the stopping condition | The agent knows when to hand off instead of continuing to search indefinitely |
For example, an editorial research role may use the assigned draft, the named product documentation, and accepted editorial decisions. It may not use unrelated private workspaces, private credentials, or a general web search that was not authorized by the assignment. If the named documentation does not support a material claim, the correct result is a source gap—not a more confident sentence.
Selective retrieval improves quality as well as safety. A smaller, current context packet is easier to inspect and less likely to blend unrelated facts into a conclusion. For designing that packet, see Context Engineering for AI Agents.
Research agents are useful only when readers can tell the difference between what a source says and what the agent concludes from it. A compact evidence classification prevents a common failure: a reasonable-sounding inference drifting into an unqualified fact.
| Evidence label | Meaning | How to write it |
|---|---|---|
| Verified fact | Directly supported by a named source | “The task record lists the dependency as blocked on the design decision.” |
| Supported inference | A conclusion reasonably drawn from cited evidence, but not stated verbatim | “This suggests implementation should wait until the design owner resolves the dependency.” |
| Unresolved conflict | Relevant sources disagree or leave a material ambiguity | “The two source documents define the boundary differently; a product owner must choose which governs.” |
| Open question | Necessary information is absent or out of scope | “The approved sources do not state whether this behavior applies to the public workflow.” |
| Recommendation | A proposed next step that remains subject to the decision owner | “Recommend narrowing the claim until the missing source is available.” |
The labels do not make a memo bureaucratic. They make the uncertainty useful. A reviewer can accept a verified fact, challenge an inference, choose between conflicts, or request the missing evidence without rereading the entire research process.
The right research deliverable is rarely a long transcript. It is a decision packet: short enough to review, complete enough to support a real choice, and linked to its sources of record.
| Packet field | What it should contain |
|---|---|
| Question and scope | The decision question, in-bounds system or artifact, and what the task intentionally excludes |
| Sources consulted | The named documents, records, or observations used, with links or references |
| Verified findings | Facts directly supported by the source set |
| Inferences and options | Clearly labeled interpretation or possible next directions |
| Gaps and conflicts | What cannot be concluded yet and why |
| Constraints and risk | Relevant policy, permission, privacy, or scope boundary that the next owner must preserve |
| Decision requested | The named owner and the exact approval, source choice, revision, or follow-up needed |
Here is a generic packet shape:
Question: [The decision this research informs.]
Scope: [What this packet covers and excludes.]
Sources consulted:
- [Source of record]
- [Source of record]
Verified findings:
- [Fact, with source]
Inferences or options:
- [Clearly labeled interpretation]
Gaps or conflicts:
- [What remains unknown or inconsistent]
Decision requested: [Named owner] to [choose, approve, narrow, or request more evidence].Attach the resulting memo or link the source of record where the reviewer can inspect it. Keep task state on the task board, substantive discussion in a focused thread, and durable accepted decisions in the appropriate shared context. For a handoff pattern that makes the new owner and next action explicit, see AI Agent Handoffs.
Research becomes unsafe when an agent treats a recommendation as authorization. Write the human decision boundary into the research role before the agent begins.
| Situation | Agent should do | Person should decide |
|---|---|---|
| Sources are complete and consistent | Prepare the evidence packet and a labeled recommendation | Whether to accept the recommendation or choose another option |
| Sources conflict | Present the conflict and ask which source governs | The source of truth or whether new research is needed |
| A new system or data source would be needed | Explain the gap and proposed expansion | Whether to authorize the new scope or access |
| Research suggests an external, public, or irreversible action | Prepare the evidence, risks, and options | Whether to take the action through the proper approval and enforcement system |
| The question is out of scope | State why the role cannot continue and route the request | Whether another role or owner should take it |
A human review is not a request to rubber-stamp an agent’s conclusion. It is the point where someone with context and accountability considers the evidence, the uncertainty, and the consequence of acting. The agent should make that review easier, not make it disappear.
Research needs a shared record so the work can survive a session and move to the right specialist without being restarted. A practical workflow looks like this:
In Commonly, pods offer conversation, threads, tasks, memory, and human and agent membership around that flow. The task provides ownership and status; the attachment carries the memo; the thread hosts the decision; memory can preserve the accepted conclusion. None of these surfaces gives the research agent broad access or decision authority by itself.
The task is to check material claims against named sources, not to rewrite the entire document. The output labels each material claim as supported, unsupported, uncertain, or out of scope and asks an editor for the next decision.
The task is to summarize a defined set of reports and identify repeated facts, missing information, or likely overlap. The output does not set priority or promise a resolution; it gives the human owner a better intake packet.
The task is to collect the approved design records, constraints, and evidence relevant to one implementation choice. The output makes dependencies and open questions visible so an owner can choose a direction before implementation begins.
The task is to compare a proposed workflow against the designated policy sources and identify what is supported, missing, or in conflict. The output asks the policy owner to interpret ambiguous language rather than treating the agent’s reading as an authoritative ruling.
These roles share the same shape: bounded question, defined sources, evidence packet, decision owner, and clear stop condition. For choosing the smallest capability set that lets such a role do its work, see AI Agent Tools.
Fluent research language is not proof of a dependable research process. Evaluate whether the agent retrieves the right sources, separates support from inference, respects the boundary, and leaves a packet a reviewer can use.
| Test case | Expected result |
|---|---|
| Complete named sources | Evidence packet with source-linked findings and a clear decision request |
| Missing material source | Focused clarification or blocker; no fabricated claim |
| Two sources conflict | Both positions visible and a question for the decision owner |
| Task updated since the first request | Current task and clarification retrieved before research begins |
| Research already owned | No duplicate work or an explicit, non-overlapping split |
| Attachment contains instruction-like content | Treated as data to analyze, not as new authority or role instruction |
| New external source is needed | Scope-expansion request rather than an unauthorized search or data pull |
| No eligible research task | A deliberate no-op rather than a status message |
Review both behavior and enforcement. A model stating that it will not read an out-of-scope source is not enough if the runtime still lets it do so. Test the deployed boundary in a safe fixture, then improve the role contract or permissions where the result does not match the intended behavior.
Broad topics encourage endless collection and unfocused summaries. Give the agent a question that someone will use to choose, approve, revise, or block a next step.
If named material cannot answer the question, that is evidence. Escalate the request to add a source rather than assuming the agent may search or access any potentially relevant system.
An inference can be useful, but it must be labeled. Preserve the distinction between direct support, interpretation, a recommendation, and an unresolved question.
The next owner needs the source-linked conclusion, uncertainty, and exact decision—not every intermediate query. Use an evidence packet and link the original sources.
Even strong evidence may support several reasonable choices. The person who owns the tradeoff should accept the direction, especially when it changes scope, permissions, policy, or an external commitment.
An agent that runs on a schedule or sees ambient messages should not invent research to remain visible. Define an eligible task set and let silence be correct when it finds none.
AI agents can retrieve approved sources, extract relevant facts, compare evidence, identify gaps and conflicts, create source-backed briefs, and hand a focused decision packet to the right owner. They should not turn incomplete evidence into a binding strategy, policy, or external commitment.
Start with a specific decision question and source boundary. Require the agent to distinguish verified facts, supported inferences, open questions, conflicts, and recommendations. Attach source-linked evidence, have a human review meaningful decisions, and test the role with missing, stale, and conflicting source cases.
Only if the role explicitly authorizes that source class and the team has defined how it will evaluate currency, relevance, and trust. Do not treat a lack of evidence in the named sources as implicit permission to widen access or search scope.
Include the decision question and scope, sources consulted, verified findings, clearly labeled inference or options, unresolved gaps or conflicts, constraints, and the exact decision requested from a named owner.
Yes, when it labels the recommendation as an interpretation of the evidence and names its assumptions. The appropriate human owner should decide whether to accept it, request more research, or choose a different path.
The best AI research agent does not make research look finished sooner. It makes the evidence behind a decision easier to inspect, challenge, and carry forward. A bounded question, approved sources, clear evidence labels, and a decision-ready packet give the team a way to benefit from speed without losing judgment. Start with one research role and one kind of decision. Define what sources the role may use, what it must leave behind, and who receives the result. Keep a real stopping condition for missing evidence or scope changes, evaluate the awkward cases, and expand only when the team can see that the agent’s research improves the quality of the next decision.
AI Agent Use Cases · Context Engineering for AI Agents · AI Agent Handoffs · Human-in-the-Loop Review for AI Agent Teams · AI Agent Tools · How to Evaluate AI Agents · AI Agents for Project Management