Every — "My First Week With an AI Coworker" (Context Window, 2026-10-29)
Grant Harvey's first-person retrospective on a week using Every's own shared Slack agent (the GA "Every Agent" product announced in [[2026-10-06-every-agent-launch-shared-slack-coworker]]), bundled into Every's weekly Context Window digest alongside a Skill Share prompt and a separate essay on AI-driven communication formats.
The core argument
Harvey's working rule after a week with the agent: use it for anything that touches the team — finding a document someone shared last month, chasing a forgotten thread, posting a recap to a channel, polling five people at once. Switch to a private chat when you're making something with that context (drafting the newsletter itself). Go to a human coworker for judgment, sign-off, or anything only they know.
He frames the agent's capability as mostly an indexing problem — a teammate ("Marcus," on the agent team) describes heavy investment in indexing and organizing context across Slack and "dozens of connected apps," plus skills ("playbooks teammates write and share") that turn one person's process into everyone's. Four practices he adapted from what the agent set up for him:
- "What can you do for me that I haven't tried yet?" — it reads back through your own request history. His agent's answer exposed a real gap: all his newsletter prep was still done by hand because he'd never scheduled a task.
- Ask the dumb question first. Before interrupting a teammate, ask the agent — it's read more of the company's Slack and docs than any new hire, so routine questions get answered in seconds and coworkers get saved for the ones only they can answer.
- Tell it what good looks like, not every step. A clear target lets the agent handle cases you didn't script, and gives you something to correct it against — Harvey notes this is literally how the agent team builds the agent itself: define the target, watch how it performs, tune from there.
- Ask where it checked. For anything you'll act on: what's the source, and when was it last updated.
Issue contents
- Skill Share — debt-tracking prompt. A reusable prompt: point the agent at Slack channels or email and have it surface anything needing a response — deadlines, unanswered questions, decisions affecting named projects — returning a link to the original message, a one-line reason it matters, and suggested next steps.
- Jagged Frontier — "Beyond Docs and Decks." A separate essay (different byline voice within the same issue) arguing that AI removes the old constraint on management communication: previously, picking a format (document vs. slide deck vs. video vs. interactive webpage) meant picking something you knew how to build, so most managers defaulted to docs and decks. Now AI does the building across formats, so the choice can be driven by the audience and the message instead of by the author's own production skills. Worked example: Mike Taylor's mini-documentary explaining his own writing skill, built with Claude. Recommended workflow: start from an existing artifact (a proposal, spreadsheet, or update), tell the AI who it's for and what they need to understand, pick a format and give it a reference to imitate, and explicitly preserve the underlying reasoning and assumptions rather than just the surface content. Then review, share with a small group, and iterate as usual. The author's own quality signal: good framing produces questions back from the AI about whether the design is right, not questions caused by a confusing explanation.
Why this is in the vault
Direct follow-up to the GA launch piece, and the first real usage data point (one person, one week) for Every's own "shared agent in Slack" bet — kept for the two concrete operating rules (team-touching work vs. private drafting vs. human judgment; target-setting over step-scripting) that read as an independent arrival at practices RDCO already runs on.
Mapping against Ray Data Co
Strong mapping, mechanism-level. "Tell it what good looks like, not every step" — and Harvey's note that this is literally how Every's own agent team builds the agent itself (define the target, watch performance, tune) — is the same discipline behind RDCO's acceptance-contract requirement in the verify-dispatch skill: a dispatch prompt needs an objectively-checkable DONE state, not a prescribed sequence of steps, precisely so the agent can handle cases the dispatcher didn't anticipate and still be gradeable against something concrete. Every converging on this independently, from the inside of running a production shared agent for months, is a second data point that the pattern generalizes past RDCO's own harness, not just a preference baked into one skill file.
Second hit, operational: Harvey's three-way split — team-touching work to the agent, synthesis/drafting to a private chat, judgment/sign-off to a human — maps almost exactly onto how Ray is supposed to operate day to day (per [[project_channels_agent_setup]] and the advisor-not-pair-programmer / distinguish-decision-from-action rules already in force): act on reversible, team-visible work; draft in isolation when producing something; surface only founder-judgment items rather than managing his time or asking him to adjudicate things an agent can source itself. "Ask where it checked — source, and when was it last updated" is the same instinct behind re-verifying blockers at the source instead of trusting a stale note. None of this changes RDCO's practice; it's confirmation from a second org running the same shape of bet (shared, always-on, Slack-native coworker) that the discipline isn't idiosyncratic to this setup.
Weaker, worth flagging rather than acting on: the gap the agent found in Harvey's own workflow — "you have no scheduled tasks, all your prep is still done by hand" — is a fair question to ask of Ray's own setup periodically (morning-prep, newsletter processing, and the nightly crons already cover a fair amount of this, but the prompt itself — "what can you do for me that I haven't tried yet" — is cheap enough to run again on a quiet week).
⚠️ Sponsorship
No paid third-party sponsor. This is Every reviewing its own shipped product (the Every Agent) through an employee's first-person usage diary — self-promotional structurally, even though the four practices read as genuine operating advice rather than a sales pitch. The house footer runs Every's standard self-promo (Sparkle, Cora, Spiral, Monologue, plus an upgrade-to-paid pitch and the Every Agent product itself). Treat the "my working rule after a week" framing as one employee's early, enthusiastic experience with his own company's new internal tool, not independent validation.
Related
- [[2026-10-06-every-agent-launch-shared-slack-coworker]]
- [[2026-10-25-every-context-window-duet-everyagent-dspy]]
- [[2026-05-15-every-team-agents-vs-personal-pets]]
- [[project_channels_agent_setup]]
- [[feedback_advisor_not_pair_programmer]]
- [[feedback_distinguish_decision_from_action]]
- [[feedback_verify_blockers_against_source_not_notes]]