Autonomous catalog enrichment for 40k products

Challenge

Skilled people stuck doing glue work

Every task touched five or more systems. Analysts spent their days copying data between tools, chasing approvals and re-checking each other's work instead of making decisions.

Objectives

What we set out to achieve

Automate the multi-step work end to end — without losing control or auditability.

• -50%Time per task
• 100%Human approval on risky steps
• 1Audit trail for every action

Strategy

Small agents, clear roles, hard limits

Rather than one "do everything" agent, we designed specialist agents — planner, researcher, executor, reviewer — each with narrow tools and explicit permissions.

Solution

An orchestrated agent team with guardrails

A graph-based orchestrator routes work between agents, calls internal tools through secure connectors, and pauses for human approval on anything irreversible.

Technology

The stack behind it

Chosen for reliability, speed and easy ownership by the client team.

• LaLangGraph
• ClClaude
• MCMCP
• PyPython
• PoPostgres
• ReRedis
• LaLangSmith
• DoDocker

Implementation

12 weeks, four phases

Weekly demos kept the client team involved at every step.

1. Weeks 1–3Process mappingTask inventory, tool access, risk review.
2. Weeks 4–7Agent buildRoles, tools, memory and orchestration graph.
3. Weeks 8–10Evaluation & guardrailsTest suites, approvals, failure handling.
4. Weeks 11–12RolloutShadow mode, then live with monitoring.

Results

Work that runs itself — safely

8x Faster, with every agent action traceable and the team focused on judgement calls instead of busywork.