Daily Paper

Published 06:00 CT · 5 stories · about 2 minutes

Claude now blocks video game cloning, and users report some reverse-engineering requests are blocked too

Agentic programming & personal agents · 2 sources, merged into one story

Bot blocked
Image: Tom's Hardware

Tom’s Hardware and Mashable report that Anthropic’s Claude has quietly started restricting some work, including game recreation, cloning and modding. Users told Tom’s Hardware that non-malicious, generic reverse-engineering and troubleshooting tasks are also being blocked.

Why it matters: If you use Claude for binary analysis or compatibility work, a refusal may now come from the model rather than from your prompt.

Sources: Tom’s Hardware (tomshardware.com) · Mashable (mashable.com)

Agents Prompt People: Using Agents for Long-Horizon Financial Work
Image: Latent Space

Arca runs ten-day agent tasks on its own harness and memory system

Arca, an AI-native wealth firm, told Latent Space that its agents do most of the background work and prompt human advisors for the final decisions. Tasks can run up to ten days, with many handoffs, so the team built an internal memory system. It owns its harness, which picks a model per task from internal benchmarks (Claude is the main provider), and a meta-agent reads advisor feedback and agent traces to suggest harness changes. Why it matters: It is a concrete design for long-running agents: persistent memory, a human approval step, and a feedback loop that tunes the harness.

latent.space · Agentic programming & personal agents

oh-my-pi 18.15.0 fixes Codex compaction and Anthropic OAuth sign-outs

oh-my-pi v18.15.0 fixes OpenAI remote compaction and Codex conversation replay, which were replaying a backend-only metadata field that Codex rejects as an unknown parameter. Anthropic OAuth accounts are no longer signed out automatically when a background usage poll’s token refresh gets invalid_grant. It also adds Modular Cloud as a provider. Why it matters: If you run omp against Codex or a rotating-token Anthropic login, these are fixes for sessions that broke on replay or logged out in the background.

github.com · oh-my-pi, Tern & Hermes

swyx: low effort plus a critique loop beats maximum effort settings

This comes from posts on X by swyx, not a published piece. He argues that a low-effort setting with a critique loop beats the highest effort levels, for people limited by throughput, and points to a babysit-runs skill in his swyxio/skills repo. In a second post he says Karpathy uses medium effort on Opus 5.5 for his latest project, and that effort level is a matter of taste. Why it matters: Effort level is a setting you control in most agent tools, so a cheaper loop with a review pass is worth testing against your own tasks.

2 sources: x.com · x.com · Agentic programming & personal agents

Below Deck season 13, episode 4 airs Monday on Bravo

Episode 4 of Below Deck season 13 airs Monday, November 2 at 8:00 p.m. Eastern, which is 7:00 p.m. Central. It airs on Bravo. Episode 3 aired on October 26. Why it matters: Bravo’s official show page lists where to watch, including streaming on Peacock.

bravotv.com · TV