Flashpoint.AIFlashpoint.AIblog
Back

Flashpoint.AI turns the real world into a function call

Connecting to humans, c. 1950
Connecting to humans, c. 1950

Until now, an agent that needed a real human's input had to guess. As of today our MCP and API are public, so it can just ask: a quick opinion poll, or a rigorous study like price elasticity of demand, conjoint, or message testing. It acts on the answers in the same run.

A few efficiencies this unlocks:

  • One interface for synthetic, yours, or ours — ground synthetic populations in real data for an instant panel, bring your own customers and users, or reach our panel partners. Global coverage.
  • Composable primitives, not fixed workflows — full survey logic, every question type. Your agent designs, branches on results, hands off to a human, resumes.
  • Measurement, not vibes — price elasticity and willingness to pay via conjoint or Van Westendorp, concept and message lift — against real buyers, before you ship.
  • Continuous, not quarterly — talk to your users every release and close the loop in the same session.
  • Priced for builders — agent-native, x402 payments, inference on your side. Easy to control costs. No platform tax on intelligence.
  • Improves with the frontier — every model release makes the agents calling us better. Nothing to upgrade on your end.

Now the part that wasn't possible before:

  • Human judgment as a function call — an agent hits a tradeoff it can't resolve, asks 50 real people, and acts on the answer in the same session. Research is a runtime dependency, not a "project."
  • Human feedback in the training loop, on demand — a model finds where it's uncertain and buys fresh human data on exactly that distribution. Active learning against reality, in real time — not batch RLHF through a vendor, weeks late.
  • Synthetic-first, escalate to real — simulate cheaply at scale, then route only the uncertain or high-stakes cases to humans. "Verify against reality" becomes a primitive, not a methodology.
  • Adaptive depth at survey scale — thousands of interviews at once, each next question chosen from the last answer. The intimacy of a 1:1, the n of a survey.
  • A self-compounding panel — an agent keeps a relationship with a cohort, remembers every answer, and always asks the next most-informative question.

Surveys are the first way to put real people in an agent's loop, in real time. More coming soon.

Read the docs to get started now.