Research

Research agents answer open questions: how a company is positioned, what the risks are in a deal, what changed in a market last quarter. The agent plans the work, reads across sources, and returns a brief with every claim cited. This page builds a due-diligence brief on one company. It continues the example from Dataset Building and Enrichment: the list found the company, the enrichment filled its record, and now one row gets a full brief.

The result

Effort medium, 65 seconds, 12 sources, output trimmed:

# Due-Diligence Brief: Qdrant (qdrant.tech)

## 1. Company background and funding history
Qdrant Solutions GmbH is a Berlin-based open-source vector-database company
... headquartered in Berlin and New York [2] with 100+ employees across 20+
countries [1] and 29,000+ GitHub stars [1] [3].

Funding history (totalling ~US$87.8M across four rounds):
- Pre-Seed: ~US$2.3M in January 2022 [4].
- Seed: ~US$3M closed April 2023, with 42 Capital among the leads [5].
- Series A: US$28M announced 23 January 2024, led by Spark Capital [6].
- Series B: US$50M announced 12 March 2026, led by AVP, with participation
  from Bosch Ventures, Unusual Ventures, Spark Capital and 42CAP [3] [4].

## 2. Leadership
- André Zayarni, CEO & Co-Founder [1].
- Andrey Vasnetsov, CTO & Co-Founder [1].
...
## 6. Legal, regulatory and reputational risks
- SOC 2 Type 2 and HIPAA certified; GDPR-aligned with a published DPA [1].
- In March 2024 IronCore Labs gave Qdrant a "weak" security maturity score
  ... later noted Qdrant added a WAF, achieved SOC 2 and launched a bug
  bounty [12].
- No publicly reported litigation or regulatory action was found. Gaps:
  Series B valuation and revenue are not disclosed.

Every [n] resolves to a claim in the trust report:

{
  "callout": 6,
  "confidence": "high",
  "reasoning": "Backed by a primary source (news)",
  "citations": [{
    "url": "https://qdrant.tech/blog/series-a-funding-round",
    "title": "Announcing Qdrant's $28M Series A Funding Round",
    "source_category": "news"
  }]
}

Note claim [5]: the seed-round figure rests on a funding-tracker page, and the report grades it high because a citation is present. That is a lead to check, not a verified fact. See what the grade is and is not.

Build it

Pick the effort

Free-text research runs at medium: a complete brief in one to three minutes for $0.10. Move to high when code consumes the result and you attach an output schema. See Effort and Cost.

Create the agent

The skill is the standing playbook: the persona, the source order, the sections, and the rule for missing facts. The goals are checkable requirements.

from nimble_python import Nimble

nimble = Nimble(api_key="YOUR-API-KEY")

agent = nimble.agents.create(
    display_name="Company Brief",
    use_case="research",
    effort="medium",
    skill=(
        "You are a research analyst writing a due-diligence brief on one named "
        "company for an investor or partner. Approach: start from the company's own "
        "website, newsroom and documentation; then funding announcements and premium "
        "press (Reuters, FT, Bloomberg, TechCrunch, Sifted); then filings where "
        "available. Cover exactly these sections in order: funding history with dates "
        "and lead investors; leadership with roles; product lines and pricing model; "
        "named customers and partners; momentum in the last 12 months; legal, "
        "regulatory and reputational risks. Cite a source for every claim. Write "
        "'Unknown' where nothing verifiable exists rather than inferring. Do not "
        "confuse the company with others of a similar name."
    ),
    goals=[
        "Every section in the skill is present, in order, even if it says Unknown",
        "Every funding round has a date, an amount and a lead investor, or says Unknown",
        "Every claim carries a citation to the page that states it",
    ],
    sources={
        "prioritize": "The company's own site and filings for facts about the company; Reuters, FT, Bloomberg, TechCrunch and Sifted for rounds and news. Funding trackers and directories are discovery only.",
    },
)

import Nimble from "@nimble-way/nimble-js";

const nimble = new Nimble({ apiKey: process.env.NIMBLE_API_KEY });

const agent = await nimble.agents.create({
  display_name: "Company Brief",
  use_case: "research",
  effort: "medium",
  skill:
    "You are a research analyst writing a due-diligence brief on one named company " +
    "for an investor or partner. Approach: start from the company's own website, " +
    "newsroom and documentation; then funding announcements and premium press " +
    "(Reuters, FT, Bloomberg, TechCrunch, Sifted); then filings where available. " +
    "Cover exactly these sections in order: funding history with dates and lead " +
    "investors; leadership with roles; product lines and pricing model; named " +
    "customers and partners; momentum in the last 12 months; legal, regulatory and " +
    "reputational risks. Cite a source for every claim. Write 'Unknown' where nothing " +
    "verifiable exists rather than inferring. Do not confuse the company with others " +
    "of a similar name.",
  goals: [
    "Every section in the skill is present, in order, even if it says Unknown",
    "Every funding round has a date, an amount and a lead investor, or says Unknown",
    "Every claim carries a citation to the page that states it",
  ],
  sources: {
    prioritize:
      "The company's own site and filings for facts about the company; Reuters, FT, " +
      "Bloomberg, TechCrunch and Sifted for rounds and news. Funding trackers and " +
      "directories are discovery only.",
  },
});

agent, err := client.Agents.New(ctx, nimble.AgentNewParams{
    DisplayName: param.NewOpt("Company Brief"),
    UseCase:     nimble.AgentNewParamsUseCaseResearch,
    Effort:      nimble.AgentNewParamsEffortMedium,
    Skill: param.NewOpt("You are a research analyst writing a due-diligence brief on one named " +
        "company for an investor or partner. Approach: start from the company's own website, " +
        "newsroom and documentation; then funding announcements and premium press (Reuters, FT, " +
        "Bloomberg, TechCrunch, Sifted); then filings where available. Cover exactly these " +
        "sections in order: funding history with dates and lead investors; leadership with " +
        "roles; product lines and pricing model; named customers and partners; momentum in the " +
        "last 12 months; legal, regulatory and reputational risks. Cite a source for every " +
        "claim. Write 'Unknown' where nothing verifiable exists rather than inferring. Do not " +
        "confuse the company with others of a similar name."),
    Goals: []string{
        "Every section in the skill is present, in order, even if it says Unknown",
        "Every funding round has a date, an amount and a lead investor, or says Unknown",
        "Every claim carries a citation to the page that states it",
    },
    Sources: nimble.AgentNewParamsSources{
        Prioritize: param.NewOpt("The company's own site and filings for facts about the company; " +
            "Reuters, FT, Bloomberg, TechCrunch and Sifted for rounds and news. Funding trackers " +
            "and directories are discovery only."),
    },
})

AGENT_ID=$(nimble agents create \
  --display-name "Company Brief" \
  --use-case research \
  --effort medium \
  --skill "You are a research analyst writing a due-diligence brief on one named company for an investor or partner. Approach: start from the company's own website, newsroom and documentation; then funding announcements and premium press (Reuters, FT, Bloomberg, TechCrunch, Sifted); then filings where available. Cover exactly these sections in order: funding history with dates and lead investors; leadership with roles; product lines and pricing model; named customers and partners; momentum in the last 12 months; legal, regulatory and reputational risks. Cite a source for every claim. Write 'Unknown' where nothing verifiable exists rather than inferring. Do not confuse the company with others of a similar name." \
  --goal "Every section in the skill is present, in order, even if it says Unknown" \
  --goal "Every funding round has a date, an amount and a lead investor, or says Unknown" \
  --goal "Every claim carries a citation to the page that states it" \
  --sources '{prioritize: "The company'"'"'s own site and filings for facts about the company; Reuters, FT, Bloomberg, TechCrunch and Sifted for rounds and news. Funding trackers and directories are discovery only."}' \
  | jq -r .id)

curl -X POST 'https://sdk.nimbleway.com/v2/agents' \
  --header 'Authorization: Bearer <YOUR-API-KEY>' \
  --header 'Content-Type: application/json' \
  --data '{
    "display_name": "Company Brief",
    "use_case": "research",
    "effort": "medium",
    "skill": "You are a research analyst writing a due-diligence brief on one named company for an investor or partner. Approach: start from the company'"'"'s own website, newsroom and documentation; then funding announcements and premium press (Reuters, FT, Bloomberg, TechCrunch, Sifted); then filings where available. Cover exactly these sections in order: funding history with dates and lead investors; leadership with roles; product lines and pricing model; named customers and partners; momentum in the last 12 months; legal, regulatory and reputational risks. Cite a source for every claim. Write Unknown where nothing verifiable exists rather than inferring. Do not confuse the company with others of a similar name.",
    "goals": [
      "Every section in the skill is present, in order, even if it says Unknown",
      "Every funding round has a date, an amount and a lead investor, or says Unknown",
      "Every claim carries a citation to the page that states it"
    ],
    "sources": {
      "prioritize": "The company'"'"'s own site and filings for facts about the company; Reuters, FT, Bloomberg, TechCrunch and Sifted for rounds and news. Funding trackers and directories are discovery only."
    }
  }'

Run it

The input names the entity with two keys (name and domain) and states the scope. Everything that applies to every company stays in the skill.

run = nimble.agents.runs.create(
    agent.id,
    input=(
        "Due-diligence brief on Qdrant (qdrant.tech, Berlin), the open-source vector "
        "database company. Cover funding history, leadership, product and pricing model, "
        "named customers and partners, momentum in the last 12 months, and legal or "
        "reputational risks. Cite every claim."
    ),
)

const run = await nimble.agents.runs.create(agent.id, {
  input:
    "Due-diligence brief on Qdrant (qdrant.tech, Berlin), the open-source vector " +
    "database company. Cover funding history, leadership, product and pricing model, " +
    "named customers and partners, momentum in the last 12 months, and legal or " +
    "reputational risks. Cite every claim.",
});

run, err := client.Agents.Runs.New(ctx, agent.ID, nimble.AgentRunNewParams{
    Input: "Due-diligence brief on Qdrant (qdrant.tech, Berlin), the open-source vector " +
        "database company. Cover funding history, leadership, product and pricing model, " +
        "named customers and partners, momentum in the last 12 months, and legal or " +
        "reputational risks. Cite every claim.",
})

RUN_ID=$(nimble agents:runs create \
  --agent-id "$AGENT_ID" \
  --input "Due-diligence brief on Qdrant (qdrant.tech, Berlin), the open-source vector database company. Cover funding history, leadership, product and pricing model, named customers and partners, momentum in the last 12 months, and legal or reputational risks. Cite every claim." \
  | jq -r .id)

curl -X POST "https://sdk.nimbleway.com/v2/agents/$AGENT_ID/runs" \
  --header 'Authorization: Bearer <YOUR-API-KEY>' \
  --header 'Content-Type: application/json' \
  --data '{
    "input": "Due-diligence brief on Qdrant (qdrant.tech, Berlin), the open-source vector database company. Cover funding history, leadership, product and pricing model, named customers and partners, momentum in the last 12 months, and legal or reputational risks. Cite every claim."
  }'

Poll while is_active is true, then fetch the result. The loop is the same for every use case: see the Quickstart.

Read the brief and check it

output.content is the brief. output.trust.claims maps every [n] to its citations. Before acting on it:

Every section present? The skill asked for six. A missing section means the agent ran out of budget or found nothing; either way it should say Unknown.
Open the citations behind the numbers. Funding amounts and dates from the company’s own announcement are solid. The same figures from a tracker or directory are leads. In the run above, claims [4] and [5] come from trackers and are graded high all the same.
Right entity? Skim the first paragraph for the domain and city you named.

Iterate

Change one thing at a time, in this order:

You see	Change
A section missing or thin	Add “even if the answer is Unknown” to the goal for that section; move the section list from `input` to `skill` if it is not there yet
Facts about a different company with a similar name	Name the entity with domain and city in `input`; add “do not confuse with…” to the skill
Trackers and directories cited as evidence	`sources.prioritize`: “trackers and directories are discovery only; cite the announcement”
Claims with no citation	Make “every claim carries a citation” a goal; it already is here
The brief must feed code	Add an output schema with one property per section and move to `high`

Best practices for research

One entity per run. Two companies in one prompt halves the depth of each and invites cross-attribution. For a comparison, run each, then a third run with previous_interaction_id to synthesize.
Free text at medium unless code reads the output. A schema turns a brief into a form to fill, and medium fills wide forms badly. With a schema, run at high and keep the schema to one property per section.
Name the entity with two keys. Name plus domain, or name plus HQ city. Similar names are the most common way a brief goes wrong.
State the sections you want, in the skill. The agent covers what it is asked to cover. A section it was not asked for is a section it may skip.
Ask for “Unknown”, and for citations per claim. Without an absence rule, gaps get filled with inference. Without a citation rule, the trust report has nothing to grade.
Tier the sources. Own site and filings first, premium press second, trackers and directories as discovery only. Say the last part explicitly. See Controlling Sources.
Do not ask for what is not public. A private company’s revenue or a seed round’s investors may not exist on the open web. The agent will say so, and the run is still billed.
x-high buys sources and length, not accuracy. Reserve it for a genuinely large corpus: years of filings, a company with heavy press. Nothing in customer pilots has shown x-high beating high on spot-checked accuracy for a brief.

From vague to production-ready

Vague
Better
Production-ready

Research Qdrant.

Which Qdrant? Research for what decision? What sections, what time window?

Write a due-diligence brief on Qdrant, the vector database company.

The entity is clearer and the purpose is stated. No sections, no absence rule, no source order.

Due-diligence brief on Qdrant (qdrant.tech, Berlin), the open-source vector
database company. Cover funding history, leadership, product and pricing
model, named customers and partners, momentum in the last 12 months, and
legal or reputational risks. Cite every claim.

With the skill above carrying the source order, section format and the Unknown rule.

Troubleshooting

The brief covers a different company

A similar name won. Put the domain and city in input, and add “do not confuse with other companies named X” to the skill. Check the first citation’s domain before reading further.

A section is missing or one line long

The agent found little, or spent its budget elsewhere. Make the section a goal (“The risks section is present even if it only says Unknown”). If the corpus really is thin, medium is not the problem and high will not fix it.

Every claim is graded high but the figures look off

Grades reflect citation presence and source type, not truth. Open the citations for the numbers that matter. Aggregators and trackers are classified as news sources and grade high. Tighten sources.prioritize to push the agent to the underlying announcement.

The run took much longer than a minute

medium typically finishes in one to three minutes; nine in ten finish under five. A run still active after ten minutes at medium is unusual: check the run id, and retry once.

I need this as JSON

Add an output_schema with one property per section (object root, strings or short arrays, every field described) and move to high. Keep it to six to eight properties. See Output Schema.

Templates

Gallery templates ship a tuned skill and goals for common research shapes. Read the copied agent’s goals and sources before running it.

Template	Built for
`competitive-intelligence`	Pricing, positioning, and feature comparison across competitors
`due-diligence`	Pre-deal company investigation: financials, leadership, legal, risk
`brand-intelligence`	Sentiment and themes across social and news, with incidents flagged
`company-profile`	Fast, sourced company one-pagers
`price-comparison`	Cross-retailer price and availability checks
`real-estate-research`	Property and market research

Next steps

You want to	Go to
Turn this brief into JSON for a pipeline	Designing the Output Schema
Run it for 300 companies	Running at Scale
Understand what the grades mean	Trust and Evaluating Results
Ask a follow-up on the same company	Follow-up runs with `previous_interaction_id`
Find the companies first	Dataset Building

Getting Started

Search Tools

Agent

Extraction Tools

Extract Template

Extract Jobs

Guides

SDKs

Admin

Proxy

The result

Build it

Best practices for research

From vague to production-ready

Troubleshooting

Templates

Next steps

​The result

​Build it

​Best practices for research

​From vague to production-ready

​Troubleshooting

​Templates

​Next steps

The result

Build it

Best practices for research

From vague to production-ready

Troubleshooting

Templates

Next steps