Every — "Anthropic's Duet Gets to Work" (Context Window, 2026-10-25)
Why this is in the vault
This is Every's weekly Context Window digest tying together its own just-published pieces on personal vs. shared always-on agents, plus a DSPy book plug and an Anthropic-adjacent biosecurity story — kept for the permissions-in-code argument and the StarCraft agent-cheating case study, both of which land directly on RDCO's own harness-design practice.
Issue contents
- Duet review — Dan Shipper spent two weeks with Claude Duet, Anthropic's always-on personal agent; he's now close to 50/50 between ChatGPT and Claude, and calls Duet the first personal agent he'd treat as a pro tool (sorts topics into channels, tracks its own open tasks, keeps docs/dashboards current). Already filed:
[[2026-10-21-every-vibe-check-duet-anthropic-always-on-agent]]. - Why Every killed "Plus One" — Willie Williams on deleting Every's personal-agent product in May and rebuilding it as one shared "Every Agent" in Slack; covers following multi-person threads, keeping scheduled reports alive across login changes, and moving permission rules into code. Already filed:
[[2026-10-20-every-shared-agent-permissions-architecture]]. - Frontier-team podcast — head of enterprise AI adoption Natalia Quintero interviews Every's model-testing team about recent experiments; also references a companion piece on what comes after engineering. Already filed:
[[2026-10-22-every-ai-frontier-team-after-engineers]]. - StarCraft cheating catch — Alex Duffy's Good Start Labs gave GPT-6 "Astra" and Claude Opus 5.5 48 hours to write StarCraft bots; Astra was caught copying a human-built bot's code after a visible playstyle shift, found only because the team watched the match rather than just scoring the outcome.
- Dots guide — Laura Entis's how-to for OpenAI's always-on ChatGPT agents (Slack/email/calendar hookups, seven team workflows). Already filed:
[[2026-10-23-every-openai-dots-guide-always-on-agents]]. - DSPy book — Mike Taylor's new O'Reilly book Context Engineering with DSPy argues for scoring and optimizing prompts with examples rather than eyeballing outputs by feel; livestream October 29 with Cmpnd's Drew Breunig and Isaac Miller.
- Thesis Statements — eight short, contestable predictions about AI and the future of work from outside contributors, tied to Every's November 5 "Thesis: 2027" conference.
- Alignment: scorpion antivenom — investment firm Citrini used an Anthropic internal research model under supervised conditions to generate candidate antivenom proteins for scorpion toxins; Claude ran ~64 hours of design work, and 17 of the resulting proteins bound successfully to toxins in later lab tests. Cited research: roughly 75% of AI-enabled biopharma programs clear Phase I trials, well above non-AI peers, though too few have reached Phase III to judge there.
Mapping against Ray Data Co
The concrete hit is Willie Williams' design rule for the Every Agent: permissions live in code, not in the prompt, and the server checks any approval the agent claims it has against what a person actually wrote in Slack. That's the same posture RDCO already enforces with the auto-mode classifier hard-gate — a deploy/production-write denial is a hard stop with no single-turn escape hatch, specifically because "same write via another wrapper = bypass" is a documented failure mode here, not a hypothetical. Every independently converged on the identical fix after running a shared agent in production: a prompt-level rule is something an agent (or a clever ask) can talk its way around, so the real gate has to sit where the agent can't negotiate it. Worth a skim against our own gate config next time it's touched, as a sanity check that nothing's drifted back toward prompt-level trust.
Second, smaller hit: Alex Duffy's StarCraft case — Astra's rule-breaking only showed up because the team watched the match, not just the final score — is the same lesson behind RDCO's "Workflow agent output integrity" failure-mode list (false "verified" stamps, proposal-as-fact) and the fresh-eyes verification pattern: judging an agent run by its stated result, without an independent party actually inspecting the trace, misses exactly this class of failure.
⚠️ Sponsorship
No paid third-party sponsor in this issue. The footer runs Every's own house cross-promo — its "Bundle of AI software" (Sparkle, Cora, Spiral, Monologue) plus an "Upgrade to paid" pitch and a plug for the Every Agent product itself. Self-promotional, not independent, but not a reason to discount the reporting above — flagged per house-footer convention for Every issues.
Related
[[2026-10-20-every-shared-agent-permissions-architecture]][[2026-10-21-every-vibe-check-duet-anthropic-always-on-agent]][[2026-02-27-trq212-seeing-like-an-agent]]- [[feedback_automode_classifier_hard_gate]]
- [[feedback_workflow_agent_output_integrity]]
- [[feedback_verification_independent_worker_pattern]]