Agents for Hire — Every (Context Window)

Why this is in the vault

Names the exact decision RDCO already made by instinct — build vs. rent vs. buy a company-wide agent — as a legible spectrum with named comparables (Shopify River, Stripe Kai, LangChain Managed Deep Agents, Notion/Lindy/Viktor), plus a second item (the Claude Code auto-mode safety study) that is directly load-bearing for how this very session is gated.

The core argument

Katie Parrott frames "company-wide agent" not as a product category but a spectrum of ownership. Shopify's River and Stripe's Kai sit at the high-ownership end: River draws on Shopify's single version-controlled repo and public Slack; Kai (built on LangChain's Deep Agents framework) took one engineer a week to ship a first version but rests on over a decade of Stripe's internal tooling, and now runs 500+ tools and 1,000+ skills. LangChain also sells the rented middle tier — Managed Deep Agents runs persistence, memory, skill loading, sandboxes, and evals while the customer supplies models, prompts, tools, and rules. Notion Custom Agents, Lindy, and Viktor sit at the buy end, trading customer control for faster time-to-start. Her actionable frame: before comparing vendors, write down where colleagues will use the agent, what company knowledge it needs, who maintains that context, and which actions require approval — those four answers determine what to build, rent, or buy.

Issue contents

Steal This Workflow — Herdr: an open-source (Apache 2.0) background server that keeps Claude Code/Codex terminal sessions alive across laptop close/disconnect, restoring tab layout and agent status on reconnect. Install via brew install herdr or curl -fsSL https://herdr.dev/install.sh | sh; 27k GitHub stars, 385k+ installs, YC-backed. Deep-fetched (herdr.dev) because it's a direct alternative/comparison point to RDCO's existing tmux + LaunchAgent session-persistence pattern for the Channels agent.
Discuss: Clay's company-wide AI writing policy (written by Sophie Alpert, shared by cofounder Varun Anand) — "More time should be spent writing a document than consuming it." Doesn't ban AI drafting; makes the human author responsible for every sentence circulated. Every applies the same standard to its own editorial guidelines.
What We're Reading: Zuckerberg's "The Future is for Everyone" memo — Meta's plan to put a personal superintelligence agent (Muse Spark, rolling into WhatsApp/Instagram/Messenger/glasses) in the hands of billions, funded by $115-135B 2026 capex, framed as safer than concentrating power in a few labs; Every's critique is that Meta would still own the apps, models, and infrastructure the "distributed" power runs through.
Links Worth a Click: Anthropic's controlled study of Claude Code auto mode — 1,053 paid professional testers, each shown one substituted dangerous permission prompt. Human reviewers caught it only 13.6% of the time (143/1,053); auto mode's classifier blocked 89% (937/1,053) of the identical commands, and auto mode blocked 800 commands humans approved vs. humans blocking only 6 auto mode allowed. Also linked: METR's blueprint for independent AI-incident investigations, OpenAI's academic-researcher usage data (heavy users ~2x as likely to submit 4+ hour tasks), and MIT's AI Agent Index (30 deployed agents scored across 45 fields). Deep-fetched the Claude Code auto-mode study (claude.com/blog) because it names the exact mechanism this session runs under.

Mapping against Ray Data Co

The auto-mode study is the sharper of the two deep-fetches: this session is running under "Auto Mode Active" right now, and feedback_automode_classifier_hard_gate already treats a denial as a hard stop rather than a retry target — Anthropic's own data (13.6% human catch rate vs. 89% classifier catch rate, worsening with session length: ~17% caught early vs. ~5% after 50+ prior prompts) is direct evidence for that policy, not just a vendor claim. It also converges with 06-reference/2026-08-10-alphasignal-claude-code-cross-session-auto-mode.md, which covered the same study from a different sender the day before — two independent newsletter pickups of one Anthropic post is a signal the finding is getting real distribution, worth citing with confidence rather than as one outlet's spin. On the build/rent/buy framing: RDCO's Ray is closer to the Kai pattern (custom-built on a general framework — Claude Code + the skills/subagent stack — with heavy proprietary context in the vault and MEMORY.md) than to a Notion/Lindy-style bought agent, which matters for calibrating expectations when the founder benchmarks Ray against "buy" competitors that start faster but can't match the depth of company-specific tooling. Herdr is a live build-vs-adopt question: RDCO's Channels agent already solves session persistence via tmux + LaunchAgent + daily 4am restart (project_channels_agent_setup) — Herdr's pitch (layout recovery, per-agent status tracking, detach-safe across laptop close) is a cleaner version of the same problem for local dev sessions specifically, not the always-on server case; worth a lightweight trial for founder-side Claude Code work on his own machine, not a replacement for the Mac Mini setup.

[[2026-08-10-alphasignal-claude-code-cross-session-auto-mode]] — same Anthropic auto-mode study (13.6% vs. 89%), covered independently by AlphaSignal the day before this issue; corroborating source.
[[2026-04-13-langchain-evals-deep-agents]] — background on LangChain's Deep Agents framework, which Stripe's Kai (profiled in this issue's SIGNAL section) is built on.
[[2026-05-09-tobi-lutke-river-public-channel-agent]] — deeper profile of Shopify's River, the other high-ownership comparable named in this issue's build/rent/buy spectrum.

⚠️ Sponsorship

No paid third-party sponsor in this issue's editorial content. The standard Every house footer promotes Every All Access (membership + $9,000+ Builder Pack credits) and Every's own product bundle (Sparkle, Cora, Spiral, Monologue) — self cross-promo only, flagged per RDCO's disclosure convention for sender self-promotion, not integrated into the SIGNAL/Discuss/Reading content above.