Working reference for the team. The build has consolidated into Made Reporting UX — Campaign Hub, the live SSO Matrix, Search Optimization, and in-app Brain Chat now run as tabs in one operator surface. Build — apps, data, agents, heartbeat. Project Mgmt — Made AI OS frame, fleet → director map, workstreams. Method — LLM-OS architecture. Infrastructure: 5 compute surfaces, 50 agents, 18 Supabase projects.
/01
Four dimensions of the build. Operator apps on top, the data layer underneath, agents running between them, and the heartbeat keeping it all coordinated. Each tile drills into its own page.
Operator surfaces. Made Reporting UX is now the primary surface — Campaign Hub (canonical funnel), the live SSO Matrix admin, Search Optimization, and Brain Chat all run here. Made Minimal (cockpit, 28 tabs) feeds in · plus this partnership dashboard.
Shared data layer. Four Made-side Supabase projects (Transfer, RAG, made-ai-os, Reporting) plus four supporting DBs. GitHub repos shipping code into all of them.
The SSO Matrix agent now resolves outreach messaging live; Brain Chat shipped in-app (2026-06-11). Plus Made-Ralph + Account Brief on the Made VPS and 3 Joe-owned in production (MADE BRAIN, RALPH, KFSH). 7 fleets underneath.
The autonomic layer. Edge functions, pg_cron jobs, Mac mini launchd, and n8n flows keeping everything pulsing. No human in the loop.
/ Recently shipped
What shipped through early June — the consolidation into Made Reporting UX.
The sso_messaging_framework (33 cells) now resolves through the made-sso-agent with a RAG self-improvement loop. Old outreach_messaging_matrix retired. Admin at /admin/messaging-framework with a Coverage Gaps panel.
13 canonical funnel statuses + a handoff lifecycle, ported from CARR (12 edge functions, MADE-PORT). One canonical outreach_messages table, one status language across Hub + Tasks.
HeyReach webhook rewritten (v4.4) after a silent payload-shape crash dropped every event; HR-truth tiles now match HeyReach 1:1 (1,804 sent / 547 accepted / 30.3%). EmailBison brought to parity — reply capture 251/251, rep handoff loop on both channels.
Full-corpus chat at /madeai/brain-chat with persistent history + a provenance panel, wrapping brain-router-v3. Next: LiteLLM proxy → router consolidation → Pydantic container (MADE_BRAIN_BUILD_PLAN.md).
New /search-optimization page under Marketing — Overview · SEO · Content · Backlinks · GEO/AI — mirroring BC's SEO/GEO harness as the made_sci tenant.
MAS + buyer-segment + the new marketing-persona layer, the canonical account-contact system, and a two-agent entity-resolution layer (Stakeholder + MADE-ID registries, propose-only HITL).
/03
The core team driving the partnership. Tasks pulled live from Hub DB — editable via Windsurf at any time. Project management hub →
Founder, BioCreative Strategies
Leads the BC side of the partnership end-to-end — the architectural counterpart to Joe’s commercial AI build. Built Made Minimal (Science Hub) from the ground up and supports all active builds across the team. Sets cadence, prioritizes projects, and brings the agent build-out to ground across every workstream.
BC operates as a teammate, not a vendor. Helps manage the overall app-build vision across departments. Supports Joe in getting custom builds over the finish line so the team can use them. Supports Greg, Lucy, Melody, Dustin individually on their workstreams. Pushes everyone’s AI capability forward through runbooks, shared tooling, and on-site work sessions. Built and maintains this dashboard as the working reference.
May 2025 → Oct 2025
Foundation era — CRM scaffolding, database build, web automation, outbound + LinkedIn ops
Nov 2025
Made Reporting UX shipped — first joint app live in production
Q1 2026
Made Minimal app, SSO matrix groundwork, weekly cadence with Joe established, operator 1:1s begin
Q2 2026
Made-Ralph entity resolution agent live, SSO matrix integrated with Joe’s positioning RAG, design port across apps
Now
Full agent-workforce alignment, this partnership dashboard, methodology codification, Tier 1–5 capability ladder
131
Total confirmed
(Apr 2025 → today)
28
Recorded in
Fireflies (~20%)
3
In-person visits
(latest May 21)
26.2h
Recorded build
session time
/04
Every team member sits on this ladder at whatever level they want. Movement is always optional. Goal is comfort + effectiveness at the rung that matches the work, not pushing anyone up the ladder unless they want it.
Uses the apps and dashboards built by Brian + Joe. No code, no DB queries. Interacts via Made Minimal, Reporting UX, documents, and team-built tools.
Greg sits here today · extended team default.
Claude Desktop and basic tools with connections to the Made AI team’s build. Read-only documentation access. Future: MCP/API connections to query data or living agents. No code writing or building.
Lucy sits here today.
Claude Desktop + MCP to Supabase, Lovable, Claude Co-Work, and web-based dev tools. Builds agents and environments following the Karpathy method. Entirely no-code but with real backend access.
Melody sits here today · Greg’s trajectory.
Windsurf, Google Antigravity, Cursor. Full database connections. Work within VPS-hosted virtual environments. Build the full agent workforce.
Dustin sits here today.
Architecture-setters. Ship custom iOS apps, run video-generation pipelines, define the design language, build agents that build agents. Self-sustaining and self-extending — the ladder doesn’t bound them.
Brian & Joe sit here today.
/05
Two pillars define how everything is designed. Every harness, agent, app, and team interaction sits on this foundation.
The Karpathy framing — Software 3.0, where the model itself is the program. The LLM is no longer just a function call; it’s a processor running an operating system you compose at runtime. We design agents the same way you’d design an OS: modular contextual systems (wikis, CLAUDE.md docs) act as the instruction layer, persistent memory is the file system, and skills are the kernel modules.
Static LLM architecture is what makes this durable. Model-agnostic routing via LiteLLM means the underlying processor swaps freely — Claude, GPT, Gemini, Qwen — without any backend change. Self-developing agents plug into skills, databases, and apps over standard interfaces (MCP, tool-calling, structured I/O), so capability grows without rewrites.
Live signal · May 28, 2026
Opus 4.8 shipped yesterday. Every agent in our stack just got an exponential upgrade with zero backend changes. Routine work also gets cheaper as models progress — every harness compounds in cost-efficiency over time. That’s the architectural payoff.
The stack you build on top of the OS. Skills are the atomic unit — markdown docs that teach a single capability. Agents compose skills into intent-driven actors with their own context pack and tool access. Harnesses orchestrate agents into repeatable pipelines: inputs → agent loop → validated outputs → side effects.
Context engineering replaces prompt engineering. Instead of crafting a clever string, you compose at invocation time: pull the right CLAUDE.md, the right skill docs, the right reference data, the right prior state — assemble, run, persist. Harness engineering makes that composition reusable. The interconnections between agents are what turn a pile of tools into a working system.
Build pipeline
skills → agents → harnesses → operating system
Different problems call for different orchestration patterns. We pick the lightest viable harness for the job and escalate when the contract demands it.
Direct Anthropic SDK loops with native tool-calling. Best when you need autonomous, multi-turn reasoning with full model context and the lowest possible latency tax.
Schema-validated agent stacks built on Pydantic AI. Strict typed I/O, automatic retries on validation failure, predictable contracts.
Declarative workflow definitions in a custom Archon YAML grammar. Composable, version-controlled, reusable across projects. The bulk of orchestrated multi-step work runs here.
Event-driven workflows for ingest, sync, and cron-style operations. Visual editor for non-engineers, full code escape hatches for edge cases. Bridges between systems where a full agent loop is overkill.
Deep dives: Project Mgmt · LLM-OS
/06
All tasks flow through these stages. Distributed across team members and tracked in Hub DB.