06-reference

technically agent access control workos

2026-10-22·reference·source: Technically·by Michael Grinich (guest, founder of WorkOS); Megan Reynolds (host, founder of infra.nyc)

"We Can Do Better Than --dangerously-skip-permissions" — Technically Speaking Ep. 1 (Michael Grinich / WorkOS)

Launch episode of Technically Speaking, a new video/podcast show from Technically (first teased in the 10/20 "What's a decision model?" issue's trailer block, now live). This is a genuinely new sub-format for the Technically relationship: a long-form (~55 min) recorded interview with show notes and timestamps, structurally distinct from both the twice-weekly "Software Eng for Vibe Coders" explainer series and the occasional single-author guest post. Host Megan Reynolds (new, founder of infra.nyc) interviews Michael Grinich, founder of WorkOS, on why AI agents break enterprise access-control models.

Why this is in the vault

Direct, named engagement with the exact problem RDCO's own harness already treats as a hard-gate design constraint — what an agent is allowed to touch, decided by code rather than by asking the model nicely — argued from the enterprise-auth side (WorkOS) rather than the harness-engineering side RDCO usually sources from.

The core argument

Grinich's framing: access control used to be a solved problem — SSO/SAML for logging in, SCIM for revoking access on exit. Agents break all three legs of that model because you can't scope access by predicting what an agent will need: ask an agent to "find ways to save the company money" and "in practice, it needs access to everything. But also nothing" — you don't know the task's actual footprint until it's mid-task.

Two specific mechanisms he names as broken:

  1. Approval pop-ups don't scale to agent speed. Human-in-the-loop confirmation works for one person making one request; it breaks once software can act thousands of times a second — his line: "to err is human, but to screw up 10,000 times per second you need a computer."
  2. Risk tolerance is company-specific, not model-specific. What's an acceptable autonomous action at Company A (WorkOS lets anything under $100 just get bought) may require sign-off at Company B — so the AI lab can't bake a universal risk policy into the model; the company has to encode its own.

His resolution: the old identity infrastructure (SSO, SAML, SCIM) doesn't get replaced for agents, it gets extended — "auth for agents and auth for AI and not just auth for people." Authentication (who are you) and authorization (what are you allowed to do) remain the two separable questions; agents just need a harder-coded authorization layer than approval dialogs, because the model can't be trusted to self-police at scale.

Mapping against Ray Data Co

Direct validation of the classifier hard-gate pattern already in the memory/playbook (feedback_automode_classifier_hard_gate): Grinich's "approval pop-ups don't scale to agent speed" is the enterprise-SaaS version of exactly why RDCO treats deploy/production-write actions as a hard stop rather than a per-call confirmation — and his "the model can't be the one to decide what's risky, the company has to" is the same reasoning Every's platform lead gave in [[2026-10-20-every-shared-agent-permissions-architecture]] for coding permission rules beneath the model instead of inside its instructions. Three independent sources (WorkOS founder, Every platform engineering, RDCO's own incident history) now converge on the same design principle: authorization has to live in a layer the model can't talk its way around.

Grinich's "the model can't be the one to decide what's risky" is also a specific application of [[verifier-as-epistemology]]'s general claim: an authorization layer only earns trust if its failure modes are independent of the thing it's checking, which is why it has to live beneath the model rather than inside its instructions. This also gives a clean conceptual label for the subject line's own --dangerously-skip-permissions reference: that Claude Code flag is the single-person, single-session version of the exact problem Grinich describes at company scale — you can turn off the approval layer for throughput, but every RDCO rule that exists today to avoid relying on it (MCP tool gating per server, the classifier hard-gate, "same write via another wrapper = bypass" in that memory) is RDCO's own small-scale answer to the same tradeoff WorkOS is selling enterprises on. Worth flagging as a gap check: RDCO's current authorization model is still instructed (AGENTS.md hard rules, memory, skill dispatch prompts), not coded-beneath-the-model the way Every's and WorkOS's are — the prompt-precedence chain explicitly says runtime system-reminders are the only mechanically-enforced layer, everything else is persuasion with no single-turn escape hatch. That's a real gap between RDCO's stated design intent and where Every/WorkOS already are.

⚠️ Sponsorship

Railway sponsors this episode of Technically Speaking (explicit pre-roll block: "Thanks to our friends at Railway for presenting this episode... Use code TECHNICALLY for $20 in free credits"). This is the show's own sponsor, not inherited from the separate Railway-sponsored "Software Eng for Vibe Coders" newsletter series — the two are distinct sponsorship relationships on distinct Technically products, confirmed by checking rather than assumed. No indication the sponsorship shaped the interview content; Railway is not mentioned inside the Grinich conversation itself, only in the surrounding newsletter copy.

Related