Geopolitical research, in English and Polish, one question at a time
Every answer comes from newsrooms you can name.
PoliticalAgent searches a fixed list of 26 publications and institutes, reads the pages it finds, and answers from that text alone. If the search or the reading fails, it tells you. It does not fill the gap.
Left of center 10
- theguardian.com
- npr.org
- vox.com
- thenation.com
- motherjones.com
- aljazeera.com
- dw.com
- brookings.edu
- carnegieendowment.org
- ecfr.eu
Center 11
- bbc.com
- bloomberg.com
- csmonitor.com
- sipri.org
- csis.org
- cfr.org
- atlanticcouncil.org
- cnas.org
- swp-berlin.org
- osw.waw.pl
- warontherocks.com
Right of center 5
- thedispatch.com
- reason.com
- foxnews.com
- aei.org
- hoover.org
Institutes in green. Placement describes the editorial balance rule, not a rating of any publication.
Three search batches · up to ten pages read per question
What happens after you press send
The first step chooses the route. Routing and conversation are free; research spends turns from your daily allowance: one for an ordinary answer, more for a deep analysis, and all of them for a forecast.
Your question is routed
A geopolitical question goes to research. A follow-up like “go deeper on the cable incidents” is answered directly, without sources. A request for a written report takes its own route, and so does an explicit request for a forecast (“give me a forecast: will …?”), which a council of seven AI models votes on.
Routing is free
The list is searched, the pages are read
Three batches of searches run across the 26 domains. Up to ten of the results are fetched and the article text is extracted — not the headline, not the search snippet.
Nothing outside the list is opened
The answer is written from that text
One model pass over what was actually read, streamed to you as it is written. If every page candidate fails to load or extract, you get an error that names the step that failed, not a confident guess.
One turn spent
A report stops and asks before it is written
Long pieces are expensive to read and worse to skim when the structure is wrong. So the outline comes first, and nothing is written until you reply.
…and that closes the researched exchange: sources read, answer written. A report can be requested after such an exchange.
Write me a report on Baltic energy security since 2022.
Here is the outline I would write. Reply to approve, revise, or cancel.
- Grid synchronisation with continental Europe
- LNG terminals and the end of piped Russian gas
- Undersea cable and pipeline incidents, 2023–2025
- What each government has committed to next
Approved, but merge three and four and add Finland.
Revisions
You can revise the outline twice. After that the report is written from the outline as it stands, so a conversation cannot loop.
When it is done
The finished report arrives with Download .md and Copy. It stays in the thread, so you can come back to it tomorrow.
Who a forecast reads
A forecast does not search the news list. It reads the published work of these 12 forecasters on their own sites, then a council of seven AI models votes on dated claims drawn from it.
Krzysztof Wojczal
Structural realism and Atlanticism; long-range forecasts of conflicts involving Poland, Russia, NATO, the United States and China
krzysztofwojczal.pl · youtube.com
Charly Salonius-Pasternak
Small-state survival, total defence, Nordic-Baltic and transatlantic security
fiia.fi
Mustafa Aydın
Middle powers; Turkey's multi-vector, transactional foreign policy
khas.edu.tr
Efraim Inbar
Military deterrence and hard-realist strategy in the Middle East; Israel
jiss.org.il
John Bew
British statecraft, diplomatic history, nuclear policy
belfercenter.org
Thomas Gomart
Great-power rivalry, geoeconomics, resources and supply chains; France and Europe
ifri.org
Yan Xuetong
Great-power theory, US-China bipolarity, moral realism
tsinghua.edu.cn
Alexander Gabuev
Eurasian security; Russia's political and economic dependence on China
carnegieendowment.org
Chaesung Chun
East Asian security, nuclear deterrence, the Korean peninsula and Seoul's balance between the United States and China
eai.or.kr
Kazuto Suzuki
Trade wars, geoeconomics, sanctions, semiconductors and resource blockades
instituteofgeoeconomics.org
Brahma Chellaney
Resource conflicts, water geopolitics, India and China
project-syndicate.org
Elizabeth Sidiropoulos
The Global South, BRICS, Africa's resources and diplomacy
saiia.org.za
What it will not do
The useful version of this list is the honest one. These are design decisions, not gaps waiting to be filled.
Invent a source
Every citation is a page that was fetched and read in that turn.
Answer without every page intact
A page the fetch could not read is dropped. Only a turn with nothing readable at all is an error — never a quieter answer. A forecast that some failed searches or its time limit cut short says so in its answer.
Reach outside the list
No blogs, no social posts, no press releases, no page you sent it. A forecast reads its own list instead: twelve forecasters' sites and one forecaster's public YouTube videos.
Forecast unless you ask
It forecasts only when you ask for a forecast. A council of seven AI models votes on dated claims from the forecasters' own writing, and a forecast spends 20 of your 50 daily research turns.
Work in English and Polish only
Questions and answers are in English or Polish; a message in any other language gets a short refusal in English. The sources are English-language publications and institutes, though a few publish some pages in Polish or German; those are read as they are and summarised in the language you asked in. One forecaster publishes in Polish; a forecast reads that work and writes its claims, the council's reasons and the analysis in the language you asked in.
Take long inputs
A question is capped at 2,000 characters, which is roughly one page.