06-reference

every agent native tool

2026-06-16·reference·source: Every·by Stella Garber

"We Built Our Own Agent-native Tool. It Overhauled How We Build Software." — @stella.f.garber

Why this is in the vault

Stella Garber (co-founder of Hoop, ex-Trello) walks through how a non-technical founding team built and iterated an internal customer-discovery analysis tool using Every's agent-native architecture principles. The piece is a ground-level case study: what broke, how they fixed it by going agent-native, and how the discipline of agent-native design bled into how they build their customer-facing product. Directly mirrors the RDCO pattern of Claude Code as always-on COO agent — this is the practitioner's version of that same architecture.

⚠️ Sponsorship

Sponsored by Microsoft Command Line — Microsoft's technical blog "written by builders, for builders." The sponsor pitch is positioned mid-article under the heading "THROW OUT THE OLD PLAYBOOK" and leads into an ad for the Command Line blog (aka.ms/CommandLineHomepage). The sponsor's framing ("the software development lifecycle wasn't designed for agentic AI") is aligned with the article's theme, so it reads as native but is disclosed via the "Want to sponsor Every?" link adjacent to it.

The core argument

Garber's startup (Hoop — an AI agent for subscription-brand churn reduction) had a classic early-stage problem: five people doing customer discovery across different tools, no shared interpretation of what prospects said. Their Monday meeting was "Brian reading from Slack notes and Granola transcripts, trying to make sense in Claude Code."

V1 — the manual transcript tool: Co-founder Justin (product background, some CS) built a Next.js + ShadCN + Supabase + Claude API tool in under 10 hours. You'd upload a Zoom transcript, it ran four or five prompts, and returned a structured analysis scored against the "PULL framework" (a Harvard Business School PMF rubric). A per-prospect summary page aggregated all calls into a relationship arc instead of single-call snapshots. The catch: everything was still manual — download transcript, upload, fill fields, wait, create a link, post to Slack. Keyword search didn't work semantically.

The agent-native pivot: Justin read Every's agent-native architecture guide and rebuilt. The design principle: instead of a hard-coded sequence of prompts in a fixed order, give the model a minimal set of tools and let it reason about how to use them. The interface moves from a destination app (where people have to go) to where people already are (Slack). The rebuilt tool needed just two tools: one to upload/read a transcript, one to add/edit a partner profile. Users dropped a transcript into Slack; the agent confirmed details, ran the analysis, and posted the summary to their feedback channel. No manual steps.

Iteration discipline: Justin ran every new feature through the agent-native architecture guidelines as a checklist — literally pasting the Every article into Claude Code and asking "where is this aligned and where is it not?" He made deliberate deviations too: LLM token-cost tracking was useful data but not something users needed to query, so he kept it outside the agent's tool surface to avoid confusion.

Payoff: By the time the newsletter was written, their Monday meeting opened with the agent's weekly synthesis — "five separate brands mentioned subscription retention as their top priority, and none of them trust existing AI tools to touch it." The insight was surfaced automatically, not extracted manually.

Article paywall note: The free version ends with Stella (the non-technical co-founder) opening Ghostty to work in the codebase herself. The paid portion covers: how she shipped an AI feature in a few hours, how the agent autonomously edited the database, and a pricing insight buried in the call data.

Mapping against Ray Data Co

Direct structural parallel. The Hoop team's discovery-call tool is nearly identical to what RDCO's Claude Code COO does across channels — ingest unstructured inputs (iMessages, Discord threads, newsletters), route through analytical prompts, synthesize and file structured outputs, surface patterns back to the founder. The agent-native principle they applied (minimal tool surface + let the model reason + bring the tool to where people already work) is exactly the design RDCO has converged on via skills + channels.

Key lift for Ray: The "use the architecture as a checklist" habit is actionable. When building or extending a skill, paste the agent-native principles and ask Claude Code "where does this deviate and is the deviation justified?" Justin's token-tracking example is the right model — deviations are fine when they're deliberate and reasoned, not accidental.

Hoop as a product watch: Hoop (hoop.app) is a direct competitor-analogue to what RDCO could productize for phData clients — an AI agent targeting subscription-brand churn. The customer discovery signals Stella describes (brands distrust AI touching retention because of off-brand responses) are a real market gap. Worth a follow.

Compound engineering lineage: This piece is the practitioner sequel to Every's earlier "compound engineering" and "how to build agent-native" articles already in the vault. The series is coalescing into a replicable method: start with three tools, maintain architecture discipline, let the agent drive the workflow rather than prescribing it.

Related