The field manual · 31 essays
Essays on proactive agents.
Adapted from internal positioning docs, postmortems, and the kind of design notes that usually live and die in PR descriptions. New ones appear roughly monthly — weather permitting.
- N°31
June 19, 2026 · 8 min
Open SWE and the architecture everyone built twice
LangChain's open-swe codifies the patterns Stripe, Ramp, and Coinbase built independently. What the most popular open-source background agent reveals.
Read the essay
- N°30
June 19, 2026 · 9 min
Open-Inspect and the Ramp Inspect blueprint
ColeMurray's background-agents project reconstructs Ramp's Inspect coding agent on Cloudflare Workers. What the most faithful open-source clone reveals.
Read the essay
- N°29
June 16, 2026 · 9 min
Tasklet skips the workflow builder
Tasklet lets the AI plan the entire workflow from a natural language prompt. It ships the richest trigger and integration ecosystem in this series, built on an architectural bet that models will keep getting better.
Read the essay
- N°28
June 16, 2026 · 5 min
Background agents and the factory model
Every major AI company ships agents that work while you sleep. The split that actually matters isn't foreground versus background — it's whether the agent waits for a ticket or goes looking for work.
Read the essay
- N°27
June 9, 2026 · 7 min
Claude Code routines and the managed runtime
Anthropic shipped routines for Claude Code — scheduled, API-triggered, and GitHub-triggered agents running on managed cloud infrastructure. It's the closest any product in this series has come to shipping all three primitives.
Read the essay
- N°26
May 29, 2026 · 8 min
Lindy gets the inbox right and the listener almost right
Lindy.ai nails email triage with push-based triggers and iMessage delivery. Beyond email, the architecture quietly falls back to polling in ways worth understanding.
Read the essay
- N°25
May 22, 2026 · 6 min
Aeon and the ceiling of cron
Aeon is a zero-infrastructure agent framework that runs 121 skills on GitHub Actions. Its self-healing architecture is genuinely novel — but the absence of a listener reveals where cron stops and proactive begins.
Read the essay
- N°24
May 20, 2026 · 6 min
MCP notifications and the listener that isn't listening yet
MCP has 9,400 servers and all the big AI companies behind it. The notification layer that would enable server-initiated push is in the spec but almost nobody uses it. The gap between what the protocol defines and what the ecosystem supports matters for anyone building proactive agents.
Read the essay
- N°23
May 20, 2026 · 6 min
Gemini Spark: what Google shipped at I/O
Google rebranded Remy to Gemini Spark at I/O 2026, shipped it on Gemini 3.5, and adopted Anthropic's MCP for third-party integrations. Here's what confirmed our predictions and what's still a black box.
Read the essay
- N°22
May 19, 2026 · 8 min
Junior.so and the AI employee question
Junior.so puts an AI in your Slack with its own email, calendar, and organizational memory. Bloomberg called it the coworker who won't stop snitching.
Read the essay
- N°21
May 19, 2026 · 8 min
Devin auto-triage and the always-on oncall
Devin's auto-triage monitors Slack, Sentry, and PagerDuty around the clock and investigates incidents on its own. Here's what the architecture reveals.
Read the essay
- N°20
May 18, 2026 · 4 min
Making Hermes proactive
The open-source Hermes agent can run cron schedules and handle webhooks, but still lacks the change detection and durable state that proactive agents need.
Read the essay
- N°19
May 18, 2026 · 6 min
Google Remy and the background agent bet
Google is testing a 24/7 Gemini-powered agent that makes purchases, sends documents, and monitors your life. Here's what the leaked architecture tells us about where proactive agents are headed.
Read the essay
- N°18
May 15, 2026 · 6 min
PostHog Code and the production signal
PostHog Code injects production data into the coding agent's context — flags, funnels, error rates. The enricher pattern points toward proactivity.
Read the essay
- N°17
May 15, 2026 · 5 min
Notion ships the proactive primitives
Notion shipped building blocks, not just an agent — Workers for developers, Custom Agents for everyone else, and an open API for external agents.
Read the essay
- N°16
May 15, 2026 · 10 min
Every tool ships an agent now
Sentry, Notion, PostHog, and CodeRabbit each shipped AI agents. Their architectures reveal a spectrum from vendor chatbot to open runtime.
Read the essay
- N°15
May 15, 2026 · 6 min
CodeRabbit's agent and the thirty-minute gap
CodeRabbit's Slack agent connects a dozen dev tools with strong automations. But real-time event detection? That's still a thirty-minute cron job.
Read the essay
- N°14
May 14, 2026 · 8 min
What forty-two percent tells us
The first realistic benchmark for proactive agents shows a 42% success ceiling. The models that win speak up least.
Read the essay
- N°13
May 13, 2026 · 7 min
What proactive agents actually cost
Every team building proactive agents hits the same wall — always-on means always spending. The ones shipping anyway found patterns worth stealing.
Read the essay
- N°12
May 12, 2026 · 9 min
A code review agent in three acts
From webhook-triggered PR reviewer to multi-surface to always-watching. Each phase taught us what proactive agents need.
Read the essay
- N°11
May 11, 2026 · 8 min
What makes proactive agents hard to build
Proactive agents look simple in demos. In production, three problems compound: knowing when to wake up, remembering across runs, and knowing when to stay quiet.
Read the essay
- N°10
May 11, 2026 · 6 min
The proactive agent wish list
A personal list of agents I want but don't have yet, across music, news, money, and work. Every entry needs the same three things.
Read the essay
- N°09
May 11, 2026 · 6 min
The prompt can't save your agent
Two pieces from the OpenClaw ecosystem show why proactive agents need a runtime, not a better system prompt.
Read the essay
- N°08
May 11, 2026 · 9 min
Where push architectures break
Push architectures break in specific, painful ways. Here's where webhooks fail and what production mitigation looks like.
Read the essay
- N°07
May 11, 2026 · 4 min
What would a magical agent actually do?
The best agents feel like an amazing intern who already handled the thing you were about to ask about. That's plumbing, not prompts.
Read the essay
- N°06
May 11, 2026 · 7 min
ChatGPT Pulse and the missing primitives
ChatGPT Pulse has one of the three primitives a proactive agent needs. Here's what's missing and what indispensable would look like.
Read the essay
- N°05
May 11, 2026 · 8 min
Building the weekly-digest agent: a production postmortem
Our weekly-digest agent scans four sources, clusters mentions, and files a GitHub issue. Here's what worked and what broke.
Read the essay
- N°04
May 10, 2026 · 7 min
Proactive agents need three primitives
A proactive agent needs three things: a clock, a listener, and an inbox, wired together with durable state. Here's why.
Read the essay
- N°03
May 10, 2026 · 10 min
The eight-week webhook tax per provider
Adding one webhook provider to a proactive agent takes a sprint. Four providers takes most of a quarter. Here's the real cost.
Read the essay
- N°02
May 10, 2026 · 5 min
The genesis of proactive agents
It started with three terminal windows and too much copy-pasting. The inbox came first, then the clock, then the listener.
Read the essay
- N°01
May 10, 2026 · 4 min
Reactive vs proactive, with examples
The same agent written twice: reactive and proactive. Same goal, same provider, different posture. The difference is who waits for whom.
Read the essay