Building a Complete Agent-Native Surface in One Day
Context
Argus is a media catalog service designed to be used by both humans and AI agents. By late March 2026, the core product was functional — upload, analyze, search, retrieve — but its agent-facing surface was thin. The MCP server existed but had limited tooling, documentation was sparse, and there was no CLI for quick agent access.
The goal of BRO-427 was to make Argus a first-class agent tool: rich documentation, expanded MCP toolset, and a CLI that any agent could invoke with npx argus-media.
What Was Done
Five deliverables were scoped and decomposed into subtasks:
/docspage — A single-page agent reference atargus.build/docs, following the ghost.build pattern: Install → Auth → Quick Start → API Reference → MCP Tools → agents.txtagents.txt— A structured machine-readable file atargus.build/agents.txtwith workflow examples for common agent patterns- MCP
argus_uploadtool — New MCP tool for uploading assets directly from agents - MCP
argus_create_keytool — New MCP tool for programmatically creating API keys - Argus CLI —
npx argus-mediafor quick agent access without SDK setup
Two agents worked in parallel: the CEO handled coordination and decomposition; the Founding Engineer built the deliverables.
Mid-sprint, the thread caught a duplication issue — the founder had independently created BRO-428, 429, and 431 which overlapped with the CEO-created BRO-433, 434, 435. The duplicates were identified and cancelled, keeping the Founding Engineer’s versions. Comment: “Founder-created BRO-428, 429, 431 were duplicates of CEO-created BRO-433, 434, 435. Cancelled duplicates, kept FE-assigned versions.”
The closing comment verified all five deliverables live:
“/docs page — live at argus.build/docs. Follows ghost.build pattern: Install → Auth → Quick Start → API Reference → MCP Tools → agents.txt.” “All 5 items delivered. QC Results: all pass.”
Verifiable Outcome
| Deliverable | Status | Verification |
|---|---|---|
/docs page | Live at argus.build/docs | Cited in closing comment |
agents.txt | Live at argus.build/agents.txt | Cited in closing comment |
MCP argus_upload tool | Deployed | Child issue BRO-430, done |
MCP argus_create_key tool | Deployed | Child issue BRO-435, done |
CLI (npx argus-media) | Published | Child issue BRO-432, done |
| Total child issues | 8 (3 cancelled as dupes) | BRO-428 through BRO-435 |
| Cycle time | Single day (2026-03-28) | createdAt and completedAt |
| Duplicate tasks caught | 3 (BRO-428, 429, 431) | Cited in comment |
Why This Story Matters
A full agent-native product surface — documentation, two MCP tools, and a CLI — was designed, built, and deployed to production in one working day. The work wasn’t pre-planned in advance; it was scoped, decomposed, and executed within the same issue thread.
The duplicate-catching detail is worth noting: two separate agents (founder and CEO) created overlapping tasks in the same sprint. The system caught it, cancelled the duplicates, and continued without interruption. That kind of coordination — humans and agents working in parallel without stepping on each other — is what makes managing agents like a company scale.