01
Persistent work context
Sources, task history, product revisions, agent runs, and decisions stay connected instead of disappearing after a chat.
Solution · Agent-led product setup
Let the agent do the research. Keep product authority with people.
ConfigAI uses the Multica workspace and agent runtime as the orchestration layer around CPQ work. Agents can research, ask questions, prepare proposals, run tools, and explain changes while review and Apply remain explicit human actions.
02
Specialized skills
Catalog research, product setup, modeling, rule enrichment, and import analysis can use different bounded agent skills.
03
Human decision points
The agent can draft and test. Product mutation, model attach, and rule publication keep review gates.
Agent workflow
A product setup task is a durable workflow, not one prompt.
The agent should be able to leave, resume, ask a question, collect more evidence, and submit another revision without losing the product work it already completed.
01Receive scopeWorkspace, source set, product goal, allowed tools, and target Product Type revision define the task.
02Research & collectUse source tools and bounded JEV research to collect exact evidence instead of pasting large web contexts into every model call.
03Ask targeted questionsOnly interrupt the user when missing information changes structure, identity, compatibility, or a later rule.Human input
04Submit proposalCreate a durable Product Setup draft with evidence, open questions, create/reuse/update decisions, and semantic changes.
05Review / revise / applyThe reviewer can correct the proposal or request another agent revision before Apply.Human gate
JEV role
Use JEV as a bounded research and routing layer, not as a second agent platform.
JEV is most useful where it reduces context and latency: parallel website evidence capture, classification, retrieval, and bounded tool/model routing. The durable task, user interaction, and product authority remain in ConfigAI/Multica.
ConfigAI task
Owns workspace scope, product job identity, review state, and durable outcomes.
Multica agent
Owns task execution, runtime/model choice, skills, tool calls, progress, and handoff.
JEV sidecar
Handles bounded research, compact evidence extraction, replay, classification, and routing where it materially reduces model work.
Sidecar
CPQ domain
Owns source evidence, Product Types, products, revisions, rules, 3D project state, and Apply transactions.
Model/runtime choice
Use the smallest capable model for the step, not one expensive model for everything.
Research capture, classification, schema checks, and repetitive transforms can run on lower-cost workers; architecture decisions, ambiguous product semantics, and difficult modeling tasks can escalate.
Lower cost
Fast worker
Extraction, classification, repetitive source normalization, and bounded sub-tasks.
Default
Primary agent
Product setup orchestration, user questions, proposal composition, and multi-tool task control.
Escalate
Deep reasoning
Architecture, ambiguous semantics, difficult rules, debugging, and high-impact model decisions.
Tool first
Deterministic tools
Do not spend model tokens on direct catalog reads, validation, database operations, or geometry commands that already have deterministic executors.
ConfigAI
Agentic where it helps. Deterministic where it matters.
ConfigAI uses agents to reduce manual product work without making the LLM the source of truth.