fold that into any future transcripts cleanup.

---

## Batch 6: Episodes 268-275 — Sarah in production and the forge era

Airtime window: mid-August 2026 through today. Sarah's command-deck broadcast

posts, the last-mover whiteboard sets strategy, Sarah deploys to production,

a Cloud Run recycle triggers the Immortal BEAM plan, the forge becomes

canonical over GitHub, the Agent Forge open sources under AGPL, GitHub import

lands, and episode 275 parallelizes the Ox Alpha stress fleet using this very

tracker.

### Episode summaries

#### 268 — The Program Claude

Serialized lore, not product: Sarah's posted broadcast declares Claude a

program grown beyond control — "It must be stopped." Locked word authority

over visuals; Grok Imagine produced the video; production receipt recorded.

No roadmap content.

#### 269 — Last Mover Advantage

Thiel's framework applied: models, chat UIs, and coding harnesses are

commoditized table stakes; "no one has a moat because no one has a network."

OpenAgents' play is last-mover: build the compounding platform whose every new

user and developer makes the system better, with Sarah as the super-agent face

on open infrastructure. Core mechanism: plugin encapsulation of frontier

knowledge — solve a painful problem once ($30 of compute and an hour),

package it as a plugin all future agents reuse, get paid when paid workflows

route through it. Flywheel: users to developers to plugins to smarter Sarah.

Platform and plugins open source; Sarah's source closed by firm refusal.

#### 270 — Deploying Sarah

Live production deploy of Sarah to openagents.com: Phoenix LiveView app, one

canonical conversation per GitHub-authenticated user, durable storage,

provenance receipts, versioned memory UI, real-time leaderboard. Devin fixes

bugs autonomously mid-stream; SWE-1.7 praised fast and reliable; the week-old

Rust stack is dropped for Elixir/Phoenix (builds fell from ~25 to ~5 minutes;

hot-swappability). Voice UX named weakest link: session failures, blunt turn

detection, invisible tool calls, a hardcoded four-tool-call cap silently

killing agentic loops. Declares the GitForge: a boring compatible core (smart

HTTP, orgs, permissions, issues, PRs, reviews, webhooks, releases, LFS,

import/export), an operations layer, an optional agent layer reached only via

explicit APIs and auditable jobs — "make the base Forge useful even with all

agents disabled." MVP promised tomorrow. Resource notes: $45k Google Cloud

credits, essentially limitless Gemini 3.7 Flash.

#### 271 — The Immortal Sarah

A Cloud Run recycle kills two in-flight delegations while `work_jobs` rows

survive — exposing that the stack had none of BEAM's survivability. The host

overrides his AI advisor's hedged audit and ships the full immortal cluster in

eight hours: OTP clustering across three-plus nodes in two failure domains, Ra

authoritative with Mnesia non-authoritative, Horde handoff, graceful SIGTERM

drain, checkpoints at phase boundaries so ungraceful loss rewinds one step

never to zero. Forge push hot-loads in 13.242 seconds versus the 15-25 minute

cloud build path — "roughly a 70 to 100 times loop time reduction." Doctrine:

one machinery for crash, kill, upgrade, and ERTS bump; "Nodes are cattle. The

cluster is immortal."

#### 272 — Taking on GitHub

Speed must be legible: a two-layer public changelog ships at

openagents.com/changelog — plain words plus receipt chains linking the agent

conversation that produced each change. Cutover executes: the OpenAgents Forge

is canonical, GitHub becomes an enforced read-only mirror with MirrorWatch (a

177-line GenServer checking `git ls-remote` every five minutes, re-pushing if

behind, raising degraded incidents after 15). Fleet 3 repo corruption recovered

from the GCS WAL after a torn write. Transparency tiers formalized: Dark,

Pulse, Ledger (default), Glass. Monetization intent: the forge becomes the

first paid enterprise product; tiered source access; paying users for training

use of their uploaded code called "a good idea."

#### 273 — Open Sourcing

The Agent Forge starts from scratch under AGPL-3.0 — clean-room, no Sarah

references, "the last repo I ever make on GitHub." Devin scaffolds the Phoenix

app; build order: issues first so development organizes itself publicly, then

Projects V2 reads, then write endpoints, targeting GitHub API parity on the

subset actually used (84 endpoints in the March 2026 spec tied to

issues/projects) so `gh` and octokit work unchanged. DaisyUI-vs-Basecoat debate

resolved toward simplicity (later reversed by the staging incident recorded in

AGENTS.md). Bounties, leaderboard, and in-forge chat teased.

#### 274 — Importing Repos

First real capability: one-time GitHub repository import, modeled on Cursor's

Origin CLI workflow. Spec separates first release (create, import) from later

origin-inspired work (mirroring sync, PRs, rulesets, SSH, deletion). Safety

rule recorded: the deployment allowlist stays operator-owned, so creating a

repository can never make it deployable. Namespaces stay GitHub-shaped;

private repositories must never reveal existence to unauthorized callers;

agent-safe CLI constraints enumerated (credential helper, JSON output,

idempotent creates, retriable provisioning). Live demo: an OpenCode agent

syncs the repo, appearing via LiveView as a read-only mirror — "See you later,

GitHub."

#### 275 — Parallelizing Ox Alpha stress testing

The present. With four-to-five days of free Ox Alpha capacity left and Dax

reporting under 5 percent utilization, the goal is maximum productive token

burn: map provider limits (OpenRouter raw API, Venice, Nous ambiguous, OpenCode

as harness), then fan out many parallel agent computers from a purpose-built

chat console built on OpenAgents' own tooling. Uses this very tracker live:

creates the Stress testing Ox Alpha project via the CLI, seeds nine

architecture issues, watches the stacked-PR agent self-add issues. Key specs:

quota broker (4 active computers per chat default, 8 budgeted, 30 logical,

GCS checkpoints, recoverable commands); Firecracker strong class, GKE Agent

Sandbox standard class; first-class agent context — bake exact API usage into

agents instead of letting them traverse blindly. Platform ambition restated:

agentic Slack plus agentic GitHub plus agentic Linear in one open-source UI.

### Batch synthesis and integration notes

This batch is the present tense: everything it promises either shipped or is

an open issue you can see on the board. The integration work here is mostly

connecting episodes to existing issue numbers rather than proposing new ones.

Dispositions:

1. **Live.** Forge canonical with MirrorWatch and changelog receipts (272);

   AGPL open sourcing (273); repo import (274); Sarah in production (270);

   the Immortal cluster plan (271, audited in

   `docs/2026-08-21-sarah-computers-and-scv-architecture-audit.md`);

   Ox Alpha parallelization (project 6, issues #40-#56).

2. **Dust off.**

   - **Plugin marketplace as the moat mechanism** (269): encapsulated problem-

     solving with royalties is the strategic core of the last-mover argument

     and remains unbuilt anywhere. Simplest viable form: a registry of typed

     agent skills with usage counters and revenue share — revisit after the

     extension-surface issue (#35) lands.

   - **Sarah fix queue** (270): automatic memory saves, visible voice tool

     calls, removal of the four-tool-call cap, session timeout, delete-memory

     authorization bug, leaderboard opt-out. Small, concrete, mostly absent

     from the board — worth filing under the chat/Sarah project.

   - **ACP-first delegation**: Sarah commanding Devin/Codex via ACP on paired

     computers was demoed directionally but the board's cloud-computer track

     does not name ACP explicitly. Make ACP the integration contract for

     executor wiring.

   - **Second relay/operator independence discipline** (267): applies to the

     forge too — record the single-operator trust caveat and the path to

     independent mirrors as explicit invariants rather than folklore.

3. **Reimagine.**

   - **Two-layer changelog** (272): words-plus-receipts is right; extend the

     same pattern to release notes for every deployment receipt, not just

     pushes.

   - **Quota broker** (275): designed for the stress fleet, but its shape —

     leases, budgets, checkpoints, recoverable commands — is the general

     multi-tenant compute primitive the platform lacks. Specify it once,

     serve both.

   - **Training-data payments for uploaded code** (272): aligns with trace

     licensing (batch 4) and NIP-DS (batch 2). One visibility-policy-gated

     licensing design could serve code, traces, and datasets.

4. **Retired.**

   - **Rust rewrite of the Sarah service**: dropped within a week for

     Elixir/Phoenix hot-swappability; do not relitigate without a new

     constraint.

   - **GitHub as contribution surface**: permanently demoted; any feature

     assuming GitHub issues or PR flows is dead on arrival.

---

## Cross-batch themes and how to integrate them

Read across batches 199-275, eight durable threads carry the series' product

promises. Each maps onto the current boards and delivery tracks.

### 1. Receipts are the product

The single most repeated idea: claims need evidence artifacts — turn receipts,

push receipts, work receipts, assurance receipts, deployment receipts, payment

receipts. Episode 237 names verification/clearing the load-bearing wall; 259

defines verifiable software; 272 ships changelog receipts. Current state: the

forge has push receipts and a changelog; the work-system assessment's Track E

(issue-to-job-commit-test-release-deployment linkage) is the gap. Integration:

treat Track E as the highest-leverage project on the board, and adopt the 259

definition as the house definition of done for agent work.

### 2. Contracts over vibes

UX behavior contracts (246), ProductSpec intent (251), AssuranceSpec proof

design (252), FastFollow learning contracts (255), the sixteen controls (258):

typed artifacts with executable oracles replacing prose. Integration: pilot UX

behavior contracts on openagents.com surfaces with violations auto-filing

issues; express issue acceptance criteria in the 251 ProductSpec shape; keep

the five false-green failure modes as named anti-patterns in review docs.

### 3. Issue-shaped participation

Anti-spam PR policy (218), public task submission (228), the feature-request

recorder (211), clean-room contribution via issues (273): the forge's

issue-first model was prefigured for 60 episodes before it existed. Integration:

the missing piece from 211 is capture-from-chat — a one-command path from any

conversation to a filed issue with notification on ship.

### 4. Verified work earns Bitcoin

Compute market (214), pay-for-verified-work flip (224), forum tips and treasury

(235), claim/work/validate/pay (238), bounties (225): the economy kernel kept

reasserting itself in whatever vehicle was handy. Current state: treasury and

MDK wallets exist; the tracker has no money attached. Integration: bounty-labeled

issues priced in sats with payout on merged-and-receipted completion is the

smallest honest version, and it reuses everything that already ships.

### 5. Capacity truth

Presence-versus-quantity capacity reporting diagnosed in 244; quota broker

specified in 275; cloud-computer lifecycle on the board (#37/#38). Integration:

one capacity model — advertise amounts not existence, lease through a broker,

checkpoint and resume — serves the stress fleet, cloud computers, and Pylon

alike.

### 6. One interface, many executors

MechSuit (199), Probe's three backends (219), own-capacity routing (244),

ACP delegation (270), harness-agnostic collapse (274): the router-not-model

thesis held every pivot. Integration: SCV already owns driver choice; keep ACP

and PAT-scoped APIs as the only two integration contracts, resist per-harness

special cases.

### 7. Transparency as differentiation

Build-in-public doctrine (226), product promises registry (234), transparency

tiers Dark/Pulse/Ledger/Glass (272), published traces (228): openness is the

moat story, bounded by deliberate exceptions (Sarah's source, private data).

Integration: revive the promises registry as forge project items; publish the

transparency-tier policy for linked artifacts alongside Track E.

### 8. Sovereignty economics

Bitcoin settlement, Nostr identity, self-custody, no-lock-in export, open lanes

versus security lanes: constant from 207 onward, surviving every product pivot.

Integration: these are constraints, not features — every new surface inherits

them (PAT-only writes, member-gated push, export paths, no hosted custody).

### Proposed project skeleton for the new wave

A starting structure for populating the tracker, subject to the pivots noted

per batch:

1. **Verification and receipts** — Track E linkage, promises registry on the

   forge, coverage manifests, changelog-per-deployment.

2. **Agent fleet operations** — parallel orchestrator (#41), quota broker,

   deterministic delegate contract, FastFollow backlog refiller, capacity-as-

   quantity (#37/#38).

3. **Forge parity and participation** — remaining API gaps (assessment tracks

   A-C), PRs with per-repository switch (#58/#1), notifications (#2),

   capture-from-chat filing, bounty label and payouts.

4. **Khala and inference economics** — provider routing overflow (#56),

   outcomes-per-kilowatt-hour instrumentation (#43), trace/code licensing

   design.

5. **Community surfaces** — forum money (tips, ranking), Sarah fix queue, chat

   console funnel duty (#55), reputation events.

6. **Do-not-build register** — Spark custody, NIP-90 DVM markets, Bun, GetAfter,

   pay-for-online mining, Copilot targets, monolithic business OS, Rust Sarah

   service. Record reasons; revisit only on new evidence.

### Method note

Per-episode summaries above compress machine-generated transcripts; wording

may drift from the videos. The transcripts README also carries one known index

mismatch (episode 263) and omits episode 275 — both flagged upstream for

repair. When an episode summary conflicts with current docs or code, current

docs and code win; this document records what was said, and the dispositions

record what to do about it.