The product should feel like senior reasoning with sources and checks, not a dashboard full of agent machinery.
Design bar
One primary action per plane.
Evidence connections must be legible.
Outputs remain reviewable drafts.
rb2 hackathon concept, product mock
From messy client intent to one verified build slice.
FieldPack makes early delivery reasoning visible: clarify intent, inspect evidence, apply tension, verify the slice, then compile one reviewable pack.
fieldpack.rb2.demo
Insurance email routing
Manual classification, policy lookup, enrichment, and department routing consume too much operator time.
IN
Clarify intentWhat changes, for whom, and why now?
readyEV
Inspect evidenceFacts, context, and recommendations stay separate.
mapped2A
Apply tensionAmbition checked against delivery realism.
draftOK
Verify sliceSmall, safe, measurable, demoable.
nextPrototype architecture: simple enough to build Friday.
The mock UI stays clean for the team, while the prototype runs one traceable path: submit a scenario, run the workflow, call specialist agents, and render the reviewed FieldPack back in the same interface.
how it works
01 · Front doorSimple UItextarea + preview
reviewed package view
reviewed package view
02 · BoundaryAPI routesubmit scenario
return package JSON
return package JSON
03 · OrchestrationMastra workflowschema-first steps
traceable state
static rb2 context pack
traceable state
static rb2 context pack
workMapAgentmaps workflow, actors, handoffs
interventionAgentfinds the safe AI slice
riskAgentmarks approvals and open questions
04 · OutputRendererMarkdown
HTML + Mermaid
back into Simple UI
HTML + Mermaid
back into Simple UI
FieldPack outcome map: what the team can trust, not just what the AI said.
A high-craft output view that combines an Obsidian-like evidence graph with a Canvas/Spark-style working board. It shows why this slice is recommended, where the proof lives, and what still needs a human decision.
data visualisation concept
AI slice
route claims
route claims
policy
lookup
lookup
mailbox
patterns
patterns
operator
handoff
handoff
safe first
build
build
unknowns
SLA
14evidence links
3open questions
1recommended slice
source clusterClient evidenceInbox examples, routing rules, policy exceptions, operator notes.
risk questionWhere should AI stop?Classification can be automated; final policy exception needs human review.
decision frameBuild sliceStart with suggested routing + confidence + explanation, not auto-send.
fieldpack artifactReviewable outcomeGraph, canvas notes, acceptance checks, and demo script in one package.
Spark-style shared workspace layer
A decision loop, not an agent inventory.
Each plane owns one question. The UI should help the team decide what is true, what is proposed, what is risky, and what is safe to build first.