01 · Challenge
Customer-success reviews often depend on subjective judgment, fragmented account context, and manual follow-up, while fully autonomous agents can create unacceptable operational risk.
Back to work
Capstone 3 · Six agents · Eight tools
AI Customer Success Manager
A Python Agents SDK capstone that coordinates six specialist agents and eight tools to assess account health, detect churn risk, recommend interventions, retain session context, and pause consequential actions for human approval.
Scroll into system
Case study brief
The problem, then the product response.
02 · Response
I built a multi-agent customer-success system that calculates account health deterministically, delegates analysis to specialist agents, remembers the session, and pauses consequential actions for approval.
System anatomy
How the system coordinates work.
The architecture is expressed as responsibilities and boundaries so the model, workflow logic, interfaces, and human controls remain inspectable.
01
Deterministic health-score calculation from account activity, adoption, sentiment, and support signals
02
Six role-specific agents coordinate analysis, planning, communication, and escalation
03
SQLite session memory and approval-gated side effects preserve context and human control
Implementation
What I built and the decisions behind it.
01
Created six specialists for health, product, support, escalation, planning, and analytics responsibilities.
02
Exposed eight tools for profile, usage, support history, product knowledge, scoring, planning, scheduling, and escalation.
03
Implemented both manager-as-tools and handoff orchestration patterns with Pydantic-validated outputs.
04
Used SQLite session memory and approval-gated side effects to preserve context without surrendering human control.
Evidence & validation
What an evaluator can inspect.
Four passing unit tests for deterministic scoring behavior.
Online and offline demonstration paths with JSON and Markdown reports.
Architecture, implementation, validation report, deck, narrated demo, and public repository.
Reflection
Agents were most useful as explainers and planners around a transparent scoring core. Deterministic numbers, structured outputs, and explicit approval made the system easier to trust and test.
Next iteration
- →Add a governed knowledge-retrieval layer.
- →Evaluate recommendation quality against expert-labelled cases.
- →Deploy role-based dashboards and production observability.
Outcomes
What the project delivered.
Six specialist agents and eight callable tools
Tested online and safe offline demonstrations
Architecture, implementation, validation report, deck, demo video, and public repository
Technology & concepts
01Python02Agents SDK03SQLite04Human Approval05Testing06Structured Outputs