Deciding which sentences are claims is the hard part
Fact-checking tools fail in one of two ways. They verify nothing important, or they demand a source for every sentence, including statements like "many teams find this frustrating." In both cases, the writer stops reading the report.
The useful work happens before any verification. You decide which statements are checkable claims, which are only frames or opinions, and which need a primary source rather than any source that repeats them. That ranking is a judgment about language. That is where an agent shows its worth.
The second criterion is weight. A hedge can stay on an unverifiable side comment. If the whole argument depends on an unverifiable claim, that claim cannot remain. You cannot see that difference from the sentence alone.
The result is three lists. The middle list holds the claims that could not be confirmed. That middle list is the deliverable.
How the agent is built in Agent Studio
The system prompt sets what counts as a claim and what counts as enough evidence. A regulated industry and a developer blog use different thresholds. That difference belongs in the prompt.
Prompt skills hold the evidence policy. The policy specifies which types of sources count as primary, how recent a source needs to be, and that a claim repeated by three secondary sources is still treated as unverified.
Retrieval comes from MCP connections. Without them, the agent could triage but not verify. That is still useful, but it is only half the product.
The agent never alters or deletes a claim. If a claim is marked as load-bearing, the flag means a human must decide before it can be published. That boundary is intentional, because if an agent quietly qualifies a sentence, it has changed your argument without telling you.
The evidence is stored, not only the verdict. After six months, the question is seldom 'did it pass' but 'what did it look at'. Only the stored evidence can answer that.
The Agent Chat node lets an ActionFlow call this agent, so verification sits inside the publishing process rather than depending on someone remembering to do it.
What this agent is built from
- System prompt: what counts as a claim and what counts as enough evidence in your field.
- Prompt skills: the evidence policy, including source recency and primary-source rules.
- MCP connections: retrieval, without which the agent can triage but not verify.
- Service and model: careful reasoning; this agent's errors are expensive and quiet.
- Agent Chat (Util): places verification inside a publishing flow rather than beside it.
- Human in the Loop: every load-bearing unverified claim requires a person's decision.
- When an ActionFlow is enough: If your claims are always the same product figures, a lookup workflow against the source of truth is more reliable than any agent.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.