← Back

AI-Assisted Delivery System

Built a named multi-agent team using Claude Code's experimental agent teams feature to support engineering and product delivery across a three-app B2B SaaS ecosystem. The system includes a PM orchestrator agent and seven specialist engineer agents, each with deep domain knowledge of their app layer, routing rules, and cross-app dependencies.

My Role

Designed the full agent architecture — defining each agent's role, domain, communication patterns, and tool access. Wrote the orchestrator and engineer agent definitions, wired Jira read-only access via Atlassian MCP, and established the shared task list and peer messaging patterns that coordinate work across the team. Applied the system to real product planning, PR review, and implementation work.

Problem

Planning and reviewing complex product work across three interconnected apps required stitching together Jira tickets, tribal knowledge, codebase context, and cross-app dependency reasoning — spread across people and tools. Code review was inconsistent and often missed architectural or migration risks. Acceptance criteria and implementation plans were written manually, creating overhead and inconsistency sprint over sprint. Overall - it was SLOW.

Approach

Designed a Claude Code agent team with a named PM orchestrator (Cartman) who reads Jira tickets via Atlassian MCP, decomposes features into vertical slices, routes work to the right specialist, and drives toward MVP. Seven engineer agents — kyle, stan, wendy, token, tweek, craig, and chef — each carry deep knowledge of their app layer (frontend, backend, DB, QA) and can communicate peer-to-peer through a shared mailbox. Agents are defined as structured markdown subagent files with explicit routing logic, anti-patterns, and output templates. The orchestrator also enforces MVP discipline: no big-bang releases, thin vertical slices, feature flags, and rolling 30/60/90 roadmap framing.

Outcome

A repeatable, context-aware delivery system that can read a Jira ticket and produce a feature brief, ticket breakdown, and engineer routing — or convene a cross-app review team for a PR. Reduced manual planning overhead, improved code review consistency, and created a foundation for AI-assisted delivery that scales with the team rather than relying on individual habits.

Stack

Claude Code, Claude Sonnet, MCP, Atlassian MCP (Jira), Filesystem integrations, Agent Teams, Cursor

Private internal project
This work was built for internal production use, so source code and detailed implementation materials are not public.