Claude’s personal agent

Anthropic has moved into the personal agent and assistant category with a product called Duet, according to Ben's Bites. Duet ships as a separate app rather than a feature inside an existing Anthropic client, runs on mobile, and is limited to the United States during a research preview phase. The newsletter's author, based outside that reach, noted he could not try it himself.
The framing matters for anyone tracking how assistant products are being packaged. Until now, most general-purpose assistants have lived inside chat interfaces or developer APIs, where the user drives each turn. A dedicated personal-agent app implies a different product surface: something that acts on a user's behalf rather than answering when asked. Anthropic has not, in this account, published details of what Duet can reach or do, so its actual capabilities remain unverified. Treat the launch as a positioning signal more than a spec sheet.
Research preview is Anthropic's usual label for an early, gated release. In practice that tends to mean invite or waitlist access, incomplete feature coverage, and behaviour that can change without notice. The newsletter's own expectation was that Duet would be competent but carry rough edges, in line with other personal-agent products. A likely trade-off for teams watching this space is capability versus trust: an agent with broad access to messages, calendars or accounts is far more useful and far harder to sandbox than a chat window. Nothing in the report describes Duet's permission model, data handling or pricing, and US-only availability means most developers cannot evaluate it directly yet.
The same issue flagged two security incidents from the prior day involving PHP and Railway. PHP is a server-side language that underpins a large share of the web, including WordPress and much custom backend code; a serious flaw there can expose sites that have not patched. Railway is a hosting platform used to deploy and run applications. The report does not name specific CVEs, affected versions or mitigations, so the practical step for teams is the ordinary one: check dependency and image versions, watch vendor advisories, and patch quickly. The newsletter author described asking Codex, OpenAI's coding agent, to scan his own projects for PHP or Railway usage, assess exposure and fix what it found — a pattern worth copying for small teams without a security function. According to the report, OpenAI helped researchers find these vulnerabilities.
On the research side, Anthropic reported that 17 of 211 protein binders designed with Claude met its experimental binding criteria against scorpion neurotoxins. That is roughly 8% of designs clearing the bar. Protein design is a domain where computational proposals are cheap and wet-lab validation is the bottleneck, so a low hit rate is not necessarily a failure; the interesting question is how the hit rate compares with prior methods, which the report does not address. The result is a concrete datapoint that model-driven design is being tested against physical assays rather than only benchmarks.
Several smaller items round out the issue. Supermemory introduced a memory graph that pulls facts out of documents, connects related memories, and handles updates by keeping earlier versions rather than overwriting them — versioned memory is a recurring pain point for agent builders whose assistants silently lose or contradict prior context. Grok Bot can build and update pages, with suggested uses such as morning briefs, design documents and weekly work summaries. AgentReady is an open specification aimed at making products usable by AI agents, structured around three capabilities: finding information, reading it, and taking action. Resend Inboxes offers a shared inbox for people and their agents, free for three inboxes. "Phone use" was highlighted as the successor to browser-use, letting agents operate real Android and iPhone devices. Ethan Mollick's book Co-Existence examines living and working alongside AI that is sometimes, but not always, better than people. On the Latent Space podcast, SF Compute CEO Evan Conrad discussed programmable data centers, GPU trading and physical settlement, and Dan Shipper argued that companies should form a frontier team to track changing models and spread what it learns internally. Ben's Bites also said that from 2027 it will work closely with four sponsor companies per quarter.
Why it matters: Developers building assistants now have another major vendor to benchmark against, though Duet's US-only preview and thin public detail mean most teams will be reading tea leaves for a while. The PHP and Railway warnings are the more immediately actionable item — if your stack touches either, audit and patch now. The protein-binder numbers and the AgentReady spec suggest two parallel tracks to watch: model output being validated by experiment, and a push to standardise how products expose themselves to agents.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
And how I run the newsletter with agents