savarin

Instruction

Run the full weekly_experiment_readout routine. Use experimental__freshness_label for all three partner sections (Suno, Coterie, HexClad). Use the production skills for everything else: readout_topline, readout_other, readout_next_steps. Execute each skill as a standalone step — do not hand-assemble any section.

Assemble the full readout in the standard order: Top-line, Suno, Coterie, HexClad, Other, Next steps & owners.

Do not reference any prior conversation or thread context. For Coterie, use this source as a one-off: https://film-framework-motor-sampling.trycloudflare.com/reports/coterie-variant-performance-evaluation

savarin

ENG-479: Weekly Experiment Readout — Index

ENG-479: Weekly Experiment Readout Quality

Outputs

1. Friday (v10) — 4-skill routine Instruction, 5 skills (orchestrator + readout_partner + readout_topline + readout_other + readout_next_steps), and output.

2. Sunday (v11) — 2-skill draft + polish

savarin

Gold Standard — Weekly Experiment Readout 2026-06-12

This is the target output for ENG-478. Cy's routine must produce output at this quality level or better with zero correction turns.

---

Weekly Experiment Readout — 2026-06-12

Top-line

• Suno has clear winners across all 3 welcome flow days, worth ~$13.5K/month — ready to scale. Results are from May 21; this week's data pipeline returned empty, so these are carry-forward numbers, not fresh evidence.

savarin

Incremental Improvement Plan: Friday → Gold Standard

Sources

• Friday output: weekly_experiment_readout v9 routine (4-skill orchestrator: readout_partner ×3, readout_other, readout_topline, readout_next_steps)
• Sunday output: weekly_readout_draft_v11 + readout_polish_v11 (2-skill pipeline)
• Gold standard: Handwritten June 12 readout
• Linear tickets: ENG-478 (training), ENG-479 (architecture exploration + content-aware optimization)
• Task folders: 20260612-1200-eng478-cold-gap-closure/, 20260614-0918-eng478-architecture-exploration/, 20260614-1500-eng479-content-aware-optimization/

savarin

Instruction

Run the weekly_readout_draft_v11 skill, then run the readout_polish_v11 skill on that output. For Coterie, use this source as a one off:

https://film-framework-motor-sampling.trycloudflare.com/reports/coterie-variant-performance-evaluation

Post both the draft and the polished version as replies in this thread.

savarin

Instruction

Run the weekly_experiment_readout routine right now as if it were Friday (not v11). Post the full output as a reply in this thread. Do not reference any prior conversation or thread for context. Please use the source below for Coterie as a one off

https://film-framework-motor-sampling.trycloudflare.com/reports/coterie-variant-performance-evaluation

savarin

Plan: Write Godot 4 PROMPT.md for Fable

Context

The Three.js browser version was visually underwhelming. Switching to Godot 4 for Vulkan rendering. Fable builds the entire game from SPEC.md + PROMPT.md. The current PROMPT.md contains Three.js instructions. SPEC.md's game design is correct but needs a Dialogue Reference extracted from bangkok.ts before the TypeScript files are deleted.

What Fable doesn't know: the Bunnag family, the Padri War, matrilineal Minangkabau culture, what a Thai noble compound looks like, who these NPCs are, why the protagonist is here, the emotional arc of the scene, the existing dialogue trees.

What Fable does know: how to structure a Godot 4 project, which nodes to use, how to write shaders, how to build UI, how to set up rendering.

savarin

← Back to Index

Building Autonomous Agent Infrastructure

The Problem

You want to deploy an AI agent that acts on behalf of your customers. It reads their Slack messages, calls their tools, sends emails, and makes decisions. But giving an agent access to real systems raises questions that tutorials don't answer.

Where do the credentials live? If the agent holds OAuth tokens for Gmail and HubSpot, what happens when the VM's memory gets snapshotted to disk? How do you let the agent call a tool without letting it call any tool? How do you give one customer's agent access to their integrations without leaking another customer's?

savarin

← Back to Index

Chapter 9: What the 8 AM Heartbeat Means

The Slack message that opened Chapter 1 traveled through every layer of infrastructure before generating a response. Each layer answered one question: what does it take to let non-deterministic software act in the world safely?

That question — implicit in every chapter — now has an answer. This final chapter assembles it.

---

savarin

← Back to Index

Chapter 8: The Data Plane

The preceding chapters examined the control plane — the gateway, provisioner, web console, and scheduler that orchestrate agent operations. But the agent itself runs inside the VM. It has its own engine, its own state, its own persona. Everything the infrastructure isolates, brokers, gates, and schedules exists to serve what lives here: a reasoning system that writes its own skills, maintains its own memory, and initiates its own actions.

This chapter works through the data plane from the inside: the AgentRuntime record that describes what boot assembled, the SOUL.md persona and its copy semantics, the skill system and how the platform-vs-agent split works, the memory store, the workspace boundaries the Apptainer sandbox enforces, the engine configuration that constrains the tool surface at startup, and the two-server topology that limits what the outside world can reach. The chapter closes by