06-reference

every shared agent permissions architecture

2026-10-20·reference·source: Every·by Willie Williams

Why We Replaced Personal Agents With One for the Whole Company -- Willie Williams (Every)

Willie Williams (Every's head of platform) walks through the engineering problems that surfaced once Every killed its personal-agent product "Plus One" (86,730 lines deleted) and consolidated onto one shared company agent, @Every, running on Claude Managed Agents. The argument: a personal agent rests on three assumptions -- it talks to one person, uses that person's logins, and can do anything that person asks -- and a shared agent breaks all three. The piece is organized around the three resulting design problems and how Every solved each:

Conversations. A Slack thread with a shared agent can run for days with multiple people talking past each other, unlike the one-to-one turn-taking of a personal agent. Every built a translation layer (engineer Paridhi Agarwal) that keeps Slack as the canonical conversation record, uses a small fast model to decide per-message whether @Every should answer, and separates unrelated requests opened in the same DM so the agent doesn't conflate them. Open question the piece admits is unresolved: when should the agent speak up unprompted.

Connections. Personal agents tie a task to one person's login; when that login expires or the person leaves, the task silently breaks and nobody may know how to fix it. Engineer Lee Knowlton's fix: tasks bind to an abstract "slot" that any valid login can occupy, so swapping credentials doesn't require re-wiring every task that depended on them.

Permissions. The hardest problem. Every encoded ownership and access rules in code beneath the model rather than in the model's instructions, specifically because "the model can't be convinced to bend rules that it shouldn't" -- in testing, an instructed-only agent tried to edit a skill it didn't own; the coded permission layer simply refused. For judgment calls outside the hard-coded rules, Every uses a single heuristic from engineer Will Gathright: how hard would this be to undo? Easy-to-undo work gets done with assumptions stated; destructive or externally-visible actions require asking first. The piece is candid about a real failure of the coded-rules approach: migrating 24 skills onto the new permissions system in July defaulted all of them to private ownership, locking out the teams that had built them together, until Lee manually widened access about a week later.

Why this is in the vault

This is a deeper, more technical companion to the Oct 6 GA-launch announcement ([[2026-10-06-every-agent-launch-shared-slack-coworker]]) -- same underlying personal-to-shared-agent pivot, different author (platform engineering lead, not the CEO) and a genuinely new argument: the specific conversation/connection/permission failure modes a shared agent hits that a personal agent never does, plus the concrete fixes (translation layer, login slots, rules-in-code, undo-cost heuristic). Worth filing separately rather than folding into the Oct 6 note.

Mapping against Ray Data Co

Strong mapping. The sharpest connection: Every's finding that permission rules must live in code beneath the model, not in instructions the model reads, because an instructed-only agent can be talked into bending its own rules -- is the same design principle behind RDCO's own prompt-precedence chain, where runtime system-reminders (mechanically enforced) outrank AGENTS.md hard rules (instructed, but explicitly "immutable from below," no single-turn escape hatch) in [[02-sops/2026-06-09-claude-md-prompt-precedence-full]]. Every's July incident -- a coded permissions migration that defaulted 24 skills to private and locked out their own builders -- is a concrete cautionary data point for RDCO before it ever hard-codes an access-control layer of its own: mechanical enforcement beats instructed enforcement, but it fails in its own way (wrong defaults, not persuasion), and that failure mode needs its own gate.

Second, the undo-cost heuristic ("how hard would this be to undo? easy → act and state assumptions; destructive or externally visible → ask first") is close to a direct restatement of two things already banked in memory: the "Complete, then ask for review" two-way-door principle and the "Auto-mode classifier hard-gate" that treats deploy/production-write actions as a hard stop. Every arrived at the same split (reversible-default-to-action vs. irreversible-default-to-ask) independently, which is useful external validation of a rule RDCO already runs on, not a new idea to adopt.

Third, the "slots" abstraction (a task binds to an abstract credential slot, not one person's login, so the agent survives personnel changes) names the exact failure mode RDCO's current single-user, single-identity (ben@raydata.co) setup doesn't need yet -- but would hit immediately if the channels-agent ever moved from solo-founder to multi-human, which is the same threshold the "Personal-license boundary" memory already flags for licensing. It's a preview of a problem, not a current one.

Fourth, the Conversations section (multi-person Slack threads, re-tagging etiquette, triage-model-per-message) mostly doesn't transfer: RDCO's channels agent is iMessage, single-human, no multi-party thread-following problem. Flagging that honestly rather than stretching the mapping -- this third of the piece is weak for RDCO specifically, even though the other two-thirds are strong.

Related

⚠️ Sponsorship

No third-party paid block. This is Every narrating its own engineering rebuild in service of its own Every Agent product (GA since Oct 6), closing with an explicit pitch section ("Why Choose Every Agent?") and the standard house footer (Every All Access subscription CTA, Sparkle/Cora/Spiral/Monologue cross-promo bundle). The candor about internal failures (the July permissions-migration lockout, the unresolved "when should the agent speak up" question) reads as unusually honest for a vendor post, but the framing throughout -- "trust Every's taste and judgment" -- still serves the pitch that customers should pay for Every's accumulated decisions rather than make their own.