06-reference

every openai dots guide always on agents

2026-10-23·reference·source: Every·by Laura Entis
dotsopenaichatgptalways-on-agentsagent-oversightevery-guide

Dots: A Guide for Putting OpenAI's Always-on Agents to Work -- Laura Entis (Every)

Why this is in the vault

A competitor-product how-to from a team running OpenAI's always-on personal agent in daily production, with four specific workflow prompts and named failure modes -- directly comparable to the founder's own self-built always-on agent and to Every's prior reporting on personal-vs-shared agent deployment shapes.

The core argument

This is the free-preview teaser for Every's paywalled how-to guide on Dot, OpenAI's proactive always-on ChatGPT agent (requires ChatGPT Business Premium). Dot monitors Slack, email, and calendar to catch errors and surface reminders, and can act directly -- answering emails, booking appointments, editing drafts. The guide distills Every's own team usage into four named workflows, each with a copyable prompt: (1) build a running list of open commitments by having Dot scan Slack/email for unanswered requests and deadlines, returning a link to the original message, why it matters, and a suggested next step; (2) catch inconsistencies across apps -- Every's own Dots have flagged an outdated article summary and a launch email referencing a feature that hadn't shipped yet; (3) set explicit rules for inbox triage and scheduling, including a named boundary -- "Don't send anything without my approval, and save any corrections I give you for next time"; (4) run multistep projects dispatched from Slack by specifying app, computer, action, and destination, with Dot reporting back a result link plus anything that needs human help. The piece is candid about failure modes: silent failures, unreliable connections, dropped tasks -- Dot needs active progress-checking, and the guide frames picking a narrow starting role with real oversight as the on-ramp. No separate curated items beyond these four workflows; they're components of one guide, not independent curated links, so this isn't a hybrid/curation note.

Mapping against Ray Data Co

Strong mapping. This piece is one of the primary sources behind [[2026-10-25-always-on-personal-agent-product-category]], the vault's concept note on the always-on-agent category and its observe-vs-act dial. The single most concrete connection: workflow #3's explicit boundary -- "Don't send anything without my approval, and save any corrections I give you for next time" -- is, word for word, the operating pattern already codified as hard rules in this agent's own configuration: no autonomous external email send ([[feedback_no_autonomous_external_email]], send-button stays human-gated) and the content-approval loop ([[feedback_content_publish_approve_then_ray_posts]]). An outside team independently arrived at the same approval-gate-plus-correction-log design for an always-on agent that RDCO has been running for months -- useful external validation that this isn't an idiosyncratic preference but a convergent pattern for anyone putting a proactive agent into daily use.

Second, workflow #2 ("catch inconsistencies across apps," with named real misses -- an outdated article summary, a launch email referencing an unshipped feature) is the exact QA role this agent already performs ad hoc, and the standing instruction to verify blockers against the source rather than stale notes ([[feedback_verify_blockers_against_source_not_notes]]) is the same discipline Every is describing as a Dot workflow rather than a human habit. Third, the guide's candid failure list -- silent failures, dropped tasks, unreliable connections requiring active oversight -- is the same "pro tool" bar Every's CEO used to judge Anthropic's Duet nine days earlier ([[2026-10-21-every-vibe-check-duet-anthropic-always-on-agent]]): a visible list of what the agent is doing and needs. The Mac Mini channels agent ([[project_channels_agent_setup]]) still lacks that first-class UI surface, relying on ops-log.md and ad hoc iMessage asks -- this is now the third independent Every piece in three weeks pointing at the same gap.

Fourth, there's a genuine tension worth naming rather than resolving: Every's own May reporting concluded personal "pet" agents don't scale at the org level in favor of one shared agent ([[2026-05-15-every-team-agents-vs-personal-pets]], proven out as the GA Every Agent in [[2026-10-06-every-agent-launch-shared-slack-coworker]]) -- yet this guide describes Every's team running individual Dots for personal commitment-tracking and inbox triage. Both can be true: a shared agent for delegated team work and a personal agent for individual oversight/triage are different jobs, not a reversal of the May thesis. Fifth, structural incentive: Every is covering a direct OpenAI competitor to the Anthropic-built Duet it reviewed glowingly days earlier and to its own Every Agent product -- the throughline across all three pieces is Every positioning itself as the authoritative guide to any always-on agent, which sells subscriptions and the Sparkle/Cora/Spiral/Monologue bundle regardless of which vendor wins.

Related

⚠️ Sponsorship

No third-party paid sponsor visible in this teaser. Standard Every house footer: "Every All Access" membership pitch plus the bundled-software cross-promo (Sparkle, Cora, Spiral, Monologue). Flagged per RDCO convention (sponsor_entity: house-promo). Separately, this is Every (which sells its own competing shared-agent product, the Every Agent, and had just published a glowing review of Anthropic's Duet) writing favorably about a third vendor's (OpenAI's) competing always-on agent -- consistent with a pattern of positioning itself as the neutral authority across all always-on-agent vendors, which is itself a commercial incentive (subscriptions, tool bundle) rather than disinterested reporting. This note is built from the free-preview teaser only; the full guide is paywalled (source_fidelity: paywall-truncated), so the four workflows above are Every's own selected highlights, not independently verified against the complete guide.