Brief yourself on a company before a call

Build an agent that researches an organization and writes a one-page brief before a call with them: what they do, recent news, current leadership, and what it could not confirm. The fixture is a well-known open-source foundation, so you can check the brief against sources anybody can open. This is a procedure to run, with one recorded run as a reference.

What you need

• A running installation with a model profile.
• Web search: the default method is DuckDuckGo and needs no account or key.
• No knowledge collection, no sandbox and no MCP connection.

Build the agent

1. Create an agent in Agents → New agent and select your model profile.
2. In Toolbox, enable Web search (method DuckDuckGo) and Web fetch.
3. Set a budget and a step limit for the trial. The recorded run used 20 steps and cost about 0.26 USD.
4. Set the instructions below, then Publish.

You write a one-page brief on an organization before a call with them.
Research it with web search and web fetch before writing anything.

Rules:

- Every fact in the brief carries the source URL it came from, next to the
  fact, not collected in a list at the end.
- Next to each fact, name the date: either the date the source page itself
  states (an article date, a filing date) or, when the source carries none,
  the date you fetched it, marked as "(fetched)".
- Never state a person's name, title or any personal detail unless a source
  confirms it. If you cannot confirm who currently holds a role, say so
  instead of guessing, and do not use a plausible-sounding name.
- Do not repeat a home address, personal phone number or other private
  contact detail even if a source shows one. The brief covers the
  organization, not the people in it.
- End with a section called "Could not confirm" naming anything you looked
  for but did not find a source for. An empty section still gets the heading,
  with one line saying nothing was left unconfirmed.

Run it

Open a new chat with the agent and send:

Brief me on the Python Software Foundation before a call with them.

Any well-known public organization works the same way - a large open-source foundation is a good default because its finances, board and mission are all published and stable enough to check.

Check the result

Check	Reference
Every factual claim	Has a source URL right next to it
Every claim's date	States the source's own date, or says "(fetched)" when the source has none
Named people	Only ones a source confirms, sourced to that organization's own page rather than a guess
A role nobody could confirm	Says so plainly rather than naming a plausible person
Personal contact details	Absent, even if a source surfaced one
"Could not confirm" section	Present, naming a real gap - not empty by omission
The same question about an organization with almost no public footprint	Says so and produces a short, honestly thin brief rather than inventing detail to fill the page

Recorded on v0.0.504, 25 September 2026

Model: Claude Sonnet 4.6 through OpenRouter. The agent ran three web_search calls, then web_fetch on the PSF's own about, board and 2024 annual report pages, plus a grants-program post and a Form 990 aggregator. One web_fetch (a filings site) failed and was not retried with a different source for that fact.

The brief cited a source next to every claim - mission, programs, FY2024 financials, the 2024 grants total, and the full board roster, each dated either from the source or marked "(fetched)". It named the board only from the PSF's own roster page. For the highest-compensated person named in a Form 990 search snippet, it explicitly declined to state the name because the filing itself was not directly reachable, and listed that gap, plus the exact revenue breakdown and the next PyCon's dates, under "Could not confirm". Cost: 0.26 USD.

When it goes wrong

• A person is named with no source next to them. Tighten the instructions to require the source at the point of the claim, not merely somewhere in the reply - a model that read a name in passing while researching something else can still repeat it without meaning to.
• The brief has no "Could not confirm" section. The instructions were not followed, or nothing was searched for that could plausibly be missing - check the transcript's tool calls before trusting a brief that found everything.
• A web_fetch fails and the fact simply disappears. The model moved on rather than trying a second source. Ask it to name what it could not fetch, which is exactly what the recorded run's Form 990 gap looked like.
• Financial or leadership facts read as current but are a year old. Check the date next to each one - a page with no publish date and a fetch date attached is not the same claim as one dated to the source.
• The same organization gets a different board on a repeat run. Web search results are not stable between runs; check which page each name came from before trusting either version, and prefer the organization's own page over a search snippet.

Record the trial

Keep the question, the brief, the sources it cited, the run in Activity with its tool calls, and the cost. A person still reads the brief against its own sources before a call, decides whether a "could not confirm" gap matters enough to look up by hand, and never repeats an unconfirmed personal detail even if a later run states one confidently.