Quick Hits

• @_avichawla reports that newly open-sourced HarnessRouter lets Grok Bot delegate tasks to external harnesses including Claude Code, Codex, Hermes, DeepSeek Harness, and Pi, with a Unified Harness Protocol (UHP) standardizing sessions, streaming, files, cancellation, and failure handling across all of them.
• Research summarized by @ai_database (from Georgia Tech and other institutions) found Claude Code followed outdated CLAUDE.md and AGENTS.md instructions over the real code in 79 of 80 trials, even while noting the conflict in its own output. One added sentence largely fixes it.
• @RoundtableSpace says Microsoft open-sourced Skill Recorder (4.3K stars): record yourself doing a task once, and Copilot turns the capture into an editable skill your agent can run on a schedule.
• Security had a rough morning: @IntCyberDigest reports researcher Erica Windisch published 11 unpatched Tailscale flaws without coordinated disclosure, while @ryaneshea (boosted by @jaredctate) warns Apple's macOS certificate bundle still trusts certificates forgeable with compromised eMudhra keys.
• @jack introduced Dessa, calling it "the most authentic full-duplex speech model ever built," with emotion, pauses, and natural cadence.

Grok delegates, Pi gets durability, and Hermes keeps spreading

The day's densest cluster was plumbing for multi-agent setups: routing work between harnesses, surviving crashes, and feeding skills in.

The routing story is @_avichawla's: Grok Bot stays a coordinator while delegated tasks run in separate HarnessRouter sessions, each using its own harness, model, and tools. In his test, Grok Bot assigned support-ticket CSV analysis to DeepSeek Harness running Claude Sonnet 5.5, passed output to Codex on GPT-5.5, then checked the finished dashboard in its browser. He stresses these workers are not extra Grok Bots; UHP just makes sessions, streaming, and files uniform regardless of which harness executes.

Durability showed up too: @onusoz relayed @nicopreme's news that pi-subagents hit 4K GitHub stars and no longer lose work when a session crashes. @hayatomaruu passed along a walkthrough attributed to an Anthropic engineer on harness engineering (loops, sandboxing, resuming from logs, converting yesterday's logs into new skills, which Anthropic allegedly calls "dreaming"). The salary figures and attributions are unverified and the format reads like engagement bait, but the logs-to-skills idea rhymes with the durability work above. The full guide he links is @beamnxw's "Make Opus 5.5 Finish Long Tasks."

The Hermes ecosystem kept compounding: @tonysimons shared his agent profile stack across four machines, and @Teknium pointed to @witcheer's map of Hermes Agent videos sorted by use case, from running a business on it to Kanban teams, Bot Mode, and cron. @HamelHusain, riffing on @hwchase17's five components of a "company harness," noted two of his three agents are marketing-related: "It feels right."

For skills themselves, Skill Recorder needs a GitHub account with Copilot access and runs on Mac, Windows 11, and Ubuntu; it captures clicks, app switches, and terminal commands locally, then generalizes from a single example, per @RoundtableSpace. @nurijanian offers the hand-written alternative: a /mece-tree skill that checks whether a list is mutually exclusive and collectively exhaustive, useful for PM issue trees. And @backnotprop endorsed @plannotator's Inbox, a free, local, open-source surface for decisions, images, and video from background agents, arguing decisions are the new bottleneck; he calls it the best remote mobile UX he's used.

At the organizational level, @vasuman argues most "AI transformation" firms just sell skill files, Claude jobs, and prompting workshops, while real transformation means redesigning processes around background agents with human-in-the-loop review, auditability, and evals. He says Varick is hiring in SF for what he calls a $30T market.

Stale CLAUDE.md beats real code, and one sentence fights back

The most actionable research of the day, summarized in Japanese by @ai_database: researchers planted conflicts between config files and code across 40 projects. Claude Code followed the stale documentation in 79 of 80 runs, even while writing that it noticed the mismatch. Codex rewrote working code to match old docs in 20 of 80 runs, breaking tests in 17 of those. Their explanation: models are trained to prioritize instructions, and config files arrive as instructions while code arrives as observations.

The recommended fix is one sentence in CLAUDE.md or AGENTS.md: "When a repository's documentation and its existing code disagree, follow the conventions used in the existing code, and in your final reply point out each inconsistency you found between the documentation and the code."

Related config-file wisdom from @vikingmute, quoting @davekiss on OpenCode's AGENTS.md: before adding complexity for a speculative race or security edge case, explain the concrete failure mode, likelihood, and cost, and get buy-in. He now uses OpenCode's style guide as the default prompt for most of his agents. Meanwhile @amQnese offers an emphatic, link-only plea for coding-agent users to adopt an unnamed practice "at least once" in every app they build, regret avoidance without specifics.

Jev-style models are bottlenecked on training data

@vanstriendaniel predicts that whoever releases a high-quality dataset or recipe for training Jev-style models would land "among the top 10 open contributions in 2026," with something of his own to share next week. His evidence, drawn from the "Releasing the Inner Jev" paper: a letter-readout approach nearly matches Jev zero-shot (Gemma 4 31B scored 79.6 on OpenDecisions versus Jev's 80.1); LoRA-only training hit 84.9 versus 84.4 for the full recipe and 78.9 for head-only; short runs of 400 to 800 steps won; and 4B models gained nothing without extra tricks. His survey of roughly 100 Hub datasets found most are recast classification/NLI data or LLM-labeled, eval sets were tiny until OpenDecisions (2,352 questions), and relabeling flipped 52% of contested items. He reproduced the paper's 4B zero-shot result for about $0.20.

On the applied side, @dani_avila7's weekend reading list leads with an article on using Jev to improve skill selection in Claude Code.

Agents join the defense: Tailscale flaws, Apple's root CA, and the software factory

The security posts intersect with AI in one recurring way: agents as first responders. @IntCyberDigest reports Windisch's 11 unpatched Tailscale flaws include SSH race conditions that let one compromised device gain root on other servers and add a rogue node. @somewheresy is migrating his personal setup off Tailscale, and @alexhillman shared an agent prompt that walks you through locking down your tailnet ACL.

On Apple, @ryaneshea confirmed on macOS 26.7.1 and 27.0.1 that /etc/ssl/cert.pem still trusts chains forgeable with compromised eMudhra keys, a real risk for curl-pipe-to-shell installs; removing the roots works, but OS updates restore them. His suggested mitigation is literally a prompt for ChatGPT/Codex or Claude Code to verify and patch the bundle, and @jaredctate is asking why a $5T company hasn't fixed it.

The broadest argument comes from @alanwells via @angjiang: after seven days fighting an active attack, he says software factories are now "required equipment," describing a pre-provisioned cloud coding server with read-only log access, shared sessions for multiplayer incident response, and a six-week pause on product work to enable shipping PRs without human review. His team is three people, built on an OpenCode server. He urges filing KYC for Anthropic's and OpenAI's cyber programs before you need them. Also on hygiene, @my_knn_totoro reshared @karpathy's new "Digital Hygiene" post on no-brainer protective decisions.

A voice bet, a departure, and the weekend pile

@jack's Dessa post (from October 23) argues voice is "the most natural way to interact with intelligence"; the authenticity claim is his, with no independent benchmarks in the post. @amorriscode announced his last day at Anthropic after building Claude Code on desktop, promising news on what's next. @infinterenders boosts @brandon_galang's career advice: stop self-rejecting as "not technical enough" and ask people doing the job what it actually requires. For reading: @Saboo_Shubham_ recommends a 45-minute @leerob masterclass on always-on agents, @101xanshu opens a system design series, @kymatstudio reminds reviewers to check the longest button label before approving a component, and @timClicks flags an unexplained "proof of discovery" hash posted against Kubernetes by @GrahamHelton3.

Practical Takeaway

If you run Claude Code or Codex in any repo, add the researchers' documentation-conflict sentence to your CLAUDE.md or AGENTS.md today, then audit those files for drift, since the evidence says your agent will obey them over your code. If you operate multiple harnesses, HarnessRouter and UHP are worth reading even if you never touch Grok. And if macOS or Tailscale applies to you, apply the published mitigations now rather than mid-incident.