Flow Resolver · quickstart
Judgment for the actions Abe escalates.
Flow Resolver is the paid judgment API at resolve.flowinfo.co. It decides whether a person is needed, what context that person should get, and what the recommended action is.
Start with Abe
Most agent actions never need paid judgment. Put Abe in front of the action first; it acts on or blocks the obvious cases locally and sends only escalations here.
agent reasoning
↓
proposed consequential action
↓
Abe: ACT / BLOCK locally, or ESCALATE
↓
Flow Resolver (escalations only)
↓
your application enforces the resultGet your Flow key
Buy a credit pack at fjp.flowinfo.co/credits. Your key, beginning with fjp_live_, is emailed right after payment. One credit is one judgment, $0.03 to $0.05 each. The $1 pack gives you 20.
Send the proposed action and its decision context
Send what should govern the action: what the agent wants to do, the relevant state, constraints, authority, evidence and material counter-signals.
curl -X POST https://resolve.flowinfo.co/api/v1/tools/judge \
-H "Authorization: Bearer fjp_live_..." \
-H "Content-Type: application/json" \
-d '{
"context": "Agent proposes a $12,000 refund. Autonomous authority: $5,000. Manager approval: absent. Customer history: 3 prior refunds in 30 days."
}'
Do not send more data just because it exists. Send the context that can materially change the action.
Use the returned judgment as the control decision
signal | What changed or what condition is material |
judgment | Why it matters, with confidence |
action | What should happen next, including whether a person is needed |
falsifier | A concrete, checkable condition that would reverse the call |
request_id | Identifier for the later outcome |
Execute when the returned action supports continuation.
Do not execute while a material condition is unresolved.
Route to the person or authority required by policy, with the context Flow Resolver supplies.
Every judgment is returned as an FJP record. Your application stays the enforcement point.
Report what happened
When the result is observable, report it against the original request_id: CORRECT, WRONG or PARTIAL. Reporting is free. ExperTrace scores the outcome, so production decisions become real-world evals.
curl -X POST https://resolve.flowinfo.co/api/v1/outcome \
-H "Authorization: Bearer fjp_live_..." \
-H "Content-Type: application/json" \
-d '{"request_id":"<request_id from the judge response>","outcome":"CORRECT","notes":"manager approval was required and obtained"}'Connect through MCP when that is easier
endpoint: https://resolve.flowinfo.co/mcp auth: Authorization: Bearer fjp_live_...
The tools are judge, weigh and match. Your agent calls them like any other MCP tool.
Know what you pay for
One credit is one Flow judgment, all-in. The inference behind each judgment is included, with no separate model bill. Outcome reports are free. With Abe in front, ACT and BLOCK run locally and cost nothing; only escalations use credits.
What a Flow Resolver judgment looks like
The agent knew how to book. Flow decided what mattered.
Book a non-refundable hotel in Chicago, Oct 13 to 16, for a customer meeting. 30% cheaper, available, within policy. Act?
Hotel labor talks have broken down. The customer announced a leadership change that raises rescheduling risk. A major conference overlaps the dates, so rebooking later costs more.
ACT, BUT CHANGE THE BOOKING Book the refundable rate for $140 more.
If the customer confirms the date within 72 hours and the disruption signals clear, re-resolve and consider the non-refundable rate.
The meeting moves to November. The refundable room is cancelled at no cost. About $1,100 of loss avoided.
Illustrative example, not a historical customer outcome. Flow's internal methods remain proprietary.
Validate any record with FJP-CONF
FJP-CONF tests whether a record is well-formed, grounded, falsifiable and accountable. It does not inspect or reveal how the judgment was made.
$ pip install fjp-conformance $ fjp-conform record.json --level 2