Atlas
Enterprise agent harness
“An agent harness that maps an entire warehouse platform, from schema to ship.”
An agentic delivery environment for a large, multi-repo enterprise warehouse-management platform. It combines .NET, MAUI, Blazor and legacy WinForms/WCF, with a knowledge vault, Graph RAG over database schemas, and evidence-before-done testing.
- Skills
- 88+
- Ticket envs
- 54
- IDEs
- 3
- Policy
- Evidence-first
CLASSIFIED · built for a client. Shown here as architecture and outcomes only: no names, no code.
The problem
The platform was a large ecosystem built up over years: several repositories, modern .NET MAUI and Blazor next to legacy WinForms and WCF, and a schema too large for anyone to hold in their head. Every change risked side effects nobody could see.
What I built
- A full-SDLC skill library. Procedural skills cover requirements, implementation, automated verification and release staging. Teams run them as slash commands.
- A ticket-centric knowledge vault. Every ticket gets its own environment: context, decisions, progress and proof.
- Graph RAG over the schemas. Before anyone touches the database, agents trace the blast radius of a change and show what it would affect.
- Three IDEs, one voice. Prompts and skills are standardized across Cursor, Claude Code and Copilot. MCP connects the agents to the issue tracker and the databases.
- Evidence-before-done. Automated E2E pipelines use Appium for regression and Maestro for fast, UI-accurate mobile flows. Nothing is “done” without a passing run.
Human in the loop
Agents propose; humans approve. Clear safety boundaries decide what an agent may do alone and what needs sign-off.
Enablement
I rolled the harness out through hands-on workshops, so cross-functional teams could use the slash commands and ship features faster, and so it wasn’t just a tool one person used.
What it flew
The same route carried very different work. Here is the range, by domain.
Warehouse floor
Handheld and desktop flows moved onto a modern mobile app and API: receiving, putaway, replenishment, paperless picking, and a post-receiving quality worklist.
▸ Each repo already had coder, test and publish skills, and a device run had to pass before the ticket could close.
Invoicing and packing
Combining invoices after they exist, parallel printing, bundle counts, gift lines, and the rules that mark work paperless.
▸ The schema and its blast radius were written down before any procedure changed.
Carriers and dispatch
A carrier API for create, track, void and close-out, plus shipping preferences and address rules that choose the carrier.
▸ SQL and the API shared one flight plan, and publish still waited for a person.
Inventory truth
Lot and expiry allocation, warehouse-versus-ERP stock gaps, and dashboard queries that were scanning too much.
▸ Agents read the real tables first, instead of guessing from the chat.
Commerce sync
Order, inventory and customer sync under deadlocks and races, including a hold so a new marketplace order was not invoiced too early.
▸ Every acceptance criterion stayed tied to the change and to evidence.
Live operations
Batch jobs, phone-order services, a live auction console, background sync, in-house health monitoring, and generated end-to-end tests the existing runner discovered on its own.
▸ Skills changed with the stack. The route, the vault and the approval gate did not.
SQL, mobile, legacy desktop, web and workers share the same stages.
The codebase and the schema are read before anyone forms a plan.
A ticket cannot close until each acceptance criterion has proof.
Risky calls and every deploy wait for an explicit yes.
Details are kept general on purpose. This was client work.