Quick Hits

• Coding agents now hold objectives instead of taking orders. Cursor shipped /goal, which gives an agent a long-lived objective to work toward "until it's fully complete," plus cloud agents that pick up work from events; @poteto says that stack helped her ship 1,000 PRs last month.
• @Benioff announced Slack Code, live with agents from Anthropic, GitHub, Cognition, and Vercel; @eltintero's review: "Slack as an ide."
• @GeneralistAI unveiled GEN-1.5, a "one-shot learner" the company says picks up new tasks in seconds, and @himanshustwts is calling it an "absolute GPT-3 moment for robotics."
• @chamath argues Databricks-style commoditization will make token consumption explode while "cost per token will crater," reacting to @alighodsi's launch of AI Extract for PDF field extraction (claimed 95% accuracy vs 87% for others, callable from SQL).
• Solo-builder corner: @elonmusk says Grok @Bot can "make an instant one-person company," @rwalk_xyz published a free CC0-licensed legal template library for startups on GitHub with an agent-friendly link, and OpenAI FDE @hi_ZachParent is polling on what deployment learnings the team should publish.

Standing objectives: Cursor's /goal and the full-autopilot playbook

Cursor's changelog, quoted by both @ManuAF6 and @poteto, describes the shift plainly: /goal gives the agent an objective it keeps working toward until it's done, and cloud agents now pick up work from events, hold a goal until it's met, and stay on course through long sessions. @poteto's post is the vivid proof of concept, though it's her own account: 1,000 shipped PRs last month, on track to double that. Her pipeline runs pstack, a personal skill library for "rigorous engineering and verification" (called /lauren-mode inside her codebase), while Grok @Bot routines farm context from Slack bug reports and X complaints to feed an "outer loop" that decides what to build next. The /goal, /loop, and /swarm commands run what she calls a Full Autopilot playbook, where agents own, verify, and ship tasks start to finish on cloud agents that keep working while she sleeps. @ManuAF6 reads the changelog as a step toward an orchestrator and eventually a "cognitive layer" in Cursor. That last part is speculation, but the shipping direction is legible.

Oversight becomes its own tooling category

The faster agents run, the more posts showed up about catching them. @gakonst says his team built Voight-Kampff, a "proof of human" gate for PR reviews, specifically to scale agent deployment without letting the security bar slip; they use it daily, and he thinks every org should build something similar, echoing @tempo's post on human authorization in agentic workflows. @WillNessAI's contribution is /code-story, a skill that turns a change into a chaptered HTML page with diffs, explainers, and diagrams, on the theory that PRs are "a terrible way to introduce a code change to someone." His deeper point is conditional: he never reads code for internal tooling, where agents prove correctness through tests and pre-merge reviews, but still reads the important parts of high-consequence services. @levie wraps the theme into an argument: experts keep the upper hand over generalists because directing, course-correcting, and judging agent output takes real domain skill, and AI hands experts more leverage than ever. He's amplifying @unclebobmartin, who describes wrestling agents as "crazy productive" but hard, focused work that draws on all his engineering skill even though he barely looks at the code.

Two bets on bring-your-own-agent: Slack Code and OpenKnowledge

@Benioff's Slack Code launch puts humans and agents in the same channel, with launch partners Anthropic, GitHub, Cognition, and Vercel, framed as "real multiplayer coding" and demoing at Dreamforce. @eltintero admits he didn't see it coming. The same day, @nickgomez announced OpenKnowledge can now chat with 35+ agents directly (the company's own post says 30+ harness integrations), enabled by the Agent Client Protocol, with Follow Mode auto-navigating the IDE to files agents are editing and Comments batching your document markup into a single feedback send. Gomez's thesis is the interesting part: people don't want a per-vendor assistant in every app, they want their own agent with its skills, MCP servers, and file-system context, and he predicts SaaS and desktop apps converge on that model. These are different products, but both are betting the surface matters less than which agent you bring to it.

Keyless search, quieter agents, and other smaller ships

@tonysimons_ flagged a quietly significant Hermes Agent update: keyless web search, meaning a fresh install can search the web with no Tavily or Exa keys and no provider setup at all. @bcherny relayed that Claude Code now has a Concise output style that leads with the result and keeps responses short. @herdrdev shipped herdr 0.8.2, with a more customizable tab bar, windows out of beta, window title sync, qwen code support, and roughly 60 fixes, after 0.8.1 was "gone before the tweet." @pidotdev resurfaced @mitsuhiko noting that Pi's new harness work is currently landing in the dev branch. And @dzhng praised an illustrations skill by @ianneo_ai as evidence that "AI slop is just a process issue," part of his "software factories (with no slop)" project.

Tally's AI-search playbook, as told by @jakezward

The most detailed business post of the day is @jakezward's breakdown of Tally's AI search strategy; every figure below is his claim, not verified reporting. He says AI search is now Tally's #1 acquisition channel, with 10,000+ new users a week discovering it through AI platforms, tracked ChatGPT registrations jumping 5x overnight in May, and growth from $258K to $422K MRR in 10 months, crossing $5M ARR with 11 people and no funding. The tactics: a comparison hub of 7 competitor pages plus 15 "best" and alternative guides structured like AI answers (rankings, "best for" labels, pricing tables, pros and cons), asking new users for the exact prompt that led them there, tracking prompt visibility against competitors, a database of articles that mention competitors but not Tally, daily clearing of Reddit and community mentions, active review gathering (4.8/5 on G2), an "AI Info" page with literal "AI assistant guidelines" for ChatGPT, Claude, Gemini, and Perplexity, an llms.txt, and a 21-tool MCP server with ChatGPT and Claude integrations. To his credit, @jakezward supplies his own caveat: ChatGPT recommends Tally because years of genuine recommendations across the internet trained it to, and the engineering came after. His follow-up post adds that the tracked discovery prompts "literally show for every prompt."

Practical Takeaway

If you're about to point agents at long-horizon work, the strongest pairing in today's posts is autonomy plus an explicit human gate. Set a /goal-style objective, but per @gakonst, put a human checkpoint on every merge, and per @WillNessAI, calibrate it by consequence: internal tooling can run on test-based verification alone, while high-stakes services deserve a real walkthrough, ideally something like a code-story page rather than a raw diff. The question worth answering before you scale: what's the smallest review gate that catches agent mistakes without becoming the bottleneck?