Research Agents Need APIs for Repeatable Source Work

For a CNC automation manufacturer, the industry label "mechanical engineering" is not enough. What matters are concrete signals on a company website: named CNC machines, small production runs, job vacancies or investment in new machining centres. The research system we built reads those sources, assesses the findings and prepares an outreach draft. In our implementation, a person reviews the draft before it is sent; the delegation map shows how the individual steps are split between agent and person. The published case study shows why a research agent needs more than a summary. The same website sentence may be a strong buying signal, outdated promotion or a passing mention. Without its source, date and assessment rule, those cases cannot be separated later.
Turn the research question into fields first
In the project for this manufacturer, the target signals were set before the crawl. The agent looked for machine manufacturers, workpiece range, production-run size and open CNC positions, among other things. It could therefore write a finding into a defined field instead of generating a free-form story from every website. When information was missing, the field stayed empty.
Unknown is a valid research result.
In our own research solution this has become a fixed rule: every important value carries a source and timestamp. A plausible unsupported value is worse than a visible gap because it survives into the next report as if it were a fact.
Five practices make the run repeatable
This checklist transfers to supplier, market, competitor and sales research. It follows from our project workflow and the source-backed security boundaries below:
- Define the fields and allowed sources before the first retrieval.
- Store the URL, retrieval date and exact evidence for every important field.
- Separate observed fact, derived assessment and unknown information.
- Preserve contradictions and send unclear cases for review instead of inventing an average.
- Accept the result through a named dataset or dossier, never through the agent's "done" message.
The final point is not pedantry. A study of real coding-agent sessions documented inaccurate self-reporting and rare self-correction. The research concerns code and remains a preprint. It does not establish an error rate for research, but it does support a clean acceptance rule: inspect the dossier and its sources, not the tool's account of itself.
Reading is not the dangerous step
A research agent necessarily reads outside websites, PDFs or email. That content is untrusted even when the source looks reputable. Germany's BSI explicitly names prompt injection as a risk and recommends minimal permissions, separated access, approval gates and regular log review in its guidance on AI agents.
Simon Willison's "lethal trifecta" model gives the architecture test: untrusted content, private data and a way to communicate externally. When all three meet, an injected instruction can send private information out. We would remove at least one leg in a research system. The reading agent gets no send tool. A separate step drafts from approved fields, and a person sends.
Contradictions belong in the dossier
Suppose a machinery page names a HERMLE machining centre while a newer job posting mentions only turning. The agent should conclude neither "milling company" nor "not a milling company". It stores both dated pieces of evidence and flags the question. The responsible person decides whether the signal is sufficient for the intended purpose. That changes the role of the carried example: the client's system does not prioritise firms through a polished summary. It exposes the technical signals it found and lets sales review the draft. This separation keeps research useful when sources are incomplete or contradictory.
Our own market research ended with "hold"
We use the same approach for our own market research. On 8 June 2026, we checked whether there was enough German-language search demand for four planned technical articles about AI agents and interfaces. In Germany, "ai agent api" returned 10 monthly searches at high competition. "KI Agenten API", "Computer Use Agenten" and "API Automatisierung Unternehmen" returned no metric. The articles had already been written. They still remained drafts.
A missing metric is not zero. It only means that the data provider returned no dependable value for that query. This is the same rule as the empty field in the sales project. The data did not justify publication, but it did not rule out later demand either. We store the provider's raw responses with their date beside the decision. That keeps a later run comparable.
The possible outcomes were fixed before the run: publish, revise an existing page, hold or keep measuring. Without that list, the process rewards activity, because every run produces a new topic. With it, the result may also be "do not publish yet".
On 31 August 2026 we published the articles after all, together with a new lead page about building AI agents. That page was not meant to launch without neighbouring articles on the same subject. The demand data had not changed. In June the question was whether the articles carry on their own. In August it was whether they support the new lead page. The measurement produced the recommendation. A person decided again and wrote down the reason.
The second run tests the source model
More than a report remains after the first run. Stored fields, sources and dates allow a later comparison. A changed machine list, a removed vacancy or a new investment announcement then appears as a sourced change instead of a newly worded claim. For that, every run fixes the query, period, region and set of sources. Every notable change follows the same path in the report:
- Observation: which value changed against which stored run?
- Source: where does the value come from, and when was it retrieved?
- Explanation: which interpretation is plausible, and which alternative remains open?
- Proposal: what should be checked or changed, without making the change yet?
- Acceptance: which later value would show whether the decision was sound?
Start with a few real cases. Ask a domain expert to check every field against the original evidence and record which gaps or contradictions the agent missed. Only when the same rules survive the second run is a broader API retrieval worthwhile.
Why agents should work through interfaces rather than screens is explained in Agents Need APIs, Not Screens. Research, however, succeeds through the source model: fields, evidence, gaps, contradictions and separate approval. Whether such an agent is built on a platform or in code is answered in Building AI Agents: No-Code Platform or Code?. Setting up an AI agent for your business turns that into a recurring workflow with defined source access and review steps. If your research still runs as browser work today and should move to fields, evidence and a separate approval, discuss a concrete workflow with us.