Wade Foster has been building Zapier as an all-remote company for fifteen years, and he opened his Running Remote 2026 session with a practical demonstration rather than a slide deck. Within the first five minutes, the room had watched him use Cursor, an AI coding agent, to run a ‘war council’ — a skill that convenes a panel of AI experts (a ruthless CFO, a contrarian board member, a customer-obsessive, and a dynamic mix of others) to rigorously critique any decision, from a hiring choice to an M&A opportunity to whether to refund a specific customer. The skill itself was built by asking the AI agent to create it.
This is the level of capability that Zapier’s team operates with daily. The presentation was less about what is possible and more about what is actually being done.
From open loops to closed loops
The most practically useful concept Foster introduced was the distinction between open-loop and closed-loop automation. Most people’s experience of AI at work involves creating more open loops — generating lists of tasks, summarising what needs to happen, producing first drafts that require human follow-through. The value is real but limited by the human bandwidth required to act on every output.
Closed-loop automation assigns tasks, tracks them, executes them, and circles back with results — without requiring a human to manage each step. Foster’s meeting follow-up pipeline, for example, analyzes his meetings and automatically drafts follow-up emails, sends Slack updates to relevant channels, and creates CRM entries. It does not produce a list of things he should do. It does the things, and shows him what it did.
The exec weekly agenda generator synthesises meeting transcripts, Slack channels, email, and chat sessions against Zapier’s strategy and goals to propose discussion topics automatically. It does not just surface information — it contextualises it against what the company is trying to accomplish.
Foster runs a weekly skill that analyses all his AI conversations, Slack, Gmail, and calendar to proactively propose new automations for eliminating repetitive work. One early output was the meeting follow-up pipeline itself — the AI noticed he was doing the same tasks after every meeting and built the system to handle them.
Governance and safety controls
The governance conversation was practical and specific. Most AI agents require all-or-nothing system access, which means a broad instruction like ‘never send emails’ might be followed 999 times and violated on the thousandth. Zapier’s approach is granular permission controls — Foster can grant his agent the ability to read Gmail and draft emails while technically preventing it from sending, ensuring that probabilistic AI behaviour is constrained by deterministic permission structures.
Foster also described a ‘how to work with me’ document that trains his personal agent to be direct, challenge his assumptions, and avoid being sycophantic. When he tested it by asking for help rebuilding Zapier’s pricing based on one customer complaint, the agent refused, noting there was insufficient evidence to justify the change. This is a designed behaviour, not a default.
For the company brain, Zapier uses monitoring skills that scan new content being added for employee PII, customer PII, or inappropriate information, either rejecting additions or alerting someone to review them.
The company brain
Three months before the conference, Zapier launched what Foster calls the company brain — a shared repository of context (strategy, values, ICPs, team charters, OKRs), connections to all business tools (Slack, Granola, Jira, CRM), and a governance layer defining permissions and audit capabilities.
The purpose is to move from individual AI acceleration — where each person develops their own workflow and the knowledge dies with them — to company-wide AI acceleration, where any skill one person builds becomes immediately available to the entire organisation. When the design team creates a slide-generating skill, anyone can use it. When marketing builds a brand writing skill, finance can access it. The hive model: knowledge flows rather than stays siloed.
This is genuinely nascent. The Zapier brain launched three months before the conference and is still being developed. But Foster was direct about why remote companies are uniquely positioned to build this: they have spent years working asynchronously, documenting SOPs, writing decisions down as a matter of course. The context AI needs to function effectively is context that remote-first organisations have been creating all along.
Driving adoption through hackathons
The single most effective driver of AI adoption that Foster identified: hackathons and hands-on building sessions. Not training programmes, not mandatory workshops, not leadership mandates — but team-based building experiences where someone slightly ahead of others teaches what they know and everyone puts hands on keyboards together.
Zapier’s AI journey evolved through three years of hackathons, from an initial code-red session focused on ChatGPT exploration three years ago to the most recent marketing hackathon where 50 marketers produced 80 projects in a single day. One of those projects: a tool for analysing ad copy against three distinct customer personas simultaneously, producing feedback in minutes that previously required hours of stakeholder coordination.
His starting point recommendation: build daily recap workflows rather than daily brief summaries. Ask your AI agent to scan meeting notes, Slack, and email not just for a list of action items, but to attempt to complete those action items — draft the emails, update the systems, surface what needs a human decision. The difference between a list and a closed loop is the difference between AI as a productivity tool and AI as a teammate.
