Good morning. Here's what's worth acting on today.
BREAKING STORY
The SBA now recognises content creators as small businesses
On 15 October the US Small Business Administration issued guidance classifying qualifying creator-led businesses under NAICS 5121 (Motion Picture and Video Industries) for SBA programme eligibility. The SBA describes it as the first federal small-business classification guidance written specifically for content creators. The SBA puts the number of creator-led businesses in that group at about 27 million, though it does not give a methodology.
The guidance covers filmmakers, podcasters, streamers, gamers, digital publishers and social-media entrepreneurs. Start with NAICS 5121, then choose the six-digit code that matches your main activity. The benefits named are SBA-backed financing, access to capital and other agency resources; no specific loan programme is listed.
The SBA also received requests for a permanent industry exception for creators and will weigh them alongside feedback on its proposed size-standards rule, which would reduce 995 size standards to 338. A Federal Register notice extended the comment period to 20 November. If you want a say in where the size line for creator businesses falls, comment before then.
My take: The paperwork caught up with reality: a solo creator is now a bankable small business, so get your books in order before you apply.
MARKET TRENDS
A familiar name nearly doubles trust in an AI answer
Raptive surveyed 3,083 US consumers in September and published the results on 14 October. With no source shown, 42.0% would trust an AI answer, and an unfamiliar source scored 41.8%. A trusted creator or expert lifted that to 75.4%, a trusted publisher to 77.2% and both together to 78.7%. 74% say knowing the source behind an AI answer is extremely or very important.
Two caveats: the figures are responses to survey source scenarios, not observed behaviour, and Raptive sells advertising for creators and publishers, so the finding suits its business.
How to make money from it: An unfamiliar citation earns no more trust than no citation, so the asset is a name readers already know. Pick one topic, publish under the same byline every time, and build a named expertise page that sets out your credentials, your best work and the questions you answer. Link every post and profile back to that page, so AI answers that cite you point to one recognisable source.
GROWTH HACK / DIY
Turn LinkedIn Advice Sessions into paid discovery calls
The idea: LinkedIn's 15 October release made paid 1:1 Advice Sessions available to Premium Business subscribers in the US. Sell a fixed-price diagnostic call that ends in a written roadmap quoting your productised offer; Brennan Dunn's Roadmapping argues that a quote inside a diagnostic report carries more weight than a cold proposal.
Do it today:
Price and list the session. Add about 20% to your hourly rate (Intro.co's guidance for expert sessions): at US$210 an hour, list 60 minutes at US$250. Sell a diagnosis and written plan to a named buyer; pin a post and a Featured link to the booking.
Send an intake, then run the call. Split 11 to 14 questions between the intake form and the call, the range Gong's analysis of 519,000 discovery calls linked to the best outcomes. On the call, quantify 3 to 4 business problems and agree which matters most; keep the prescription for the roadmap.
Deliver the roadmap within 24 hours. One to two pages: the problems, your approach and your fixed-price productised offer, less the session fee if they accept within 14 days.
What you'll need: LinkedIn Premium Business at US$69.99 a month (as of 18 October 2026), verified profiles for you and the booker, a fixed-price offer, an intake form and a roadmap template; setup takes three hours, each session two. Engadget reports LinkedIn takes no cut; LinkedIn has not confirmed fee, payout or refund terms, so check the Advice Sessions Help page. Outside the US, use Topmate or Calendly plus Stripe.
AGENT LAB
Approval gates: make your agent ask before it emails, pays or posts
An approval gate sits between an agent deciding to act and the action happening. The agent drafts the call, pauses, shows you the tool name and arguments, and waits for you to approve, edit, reject or respond. Use it on side effects you cannot reverse: client emails, payments, public posts, deletions. Leave read-only tools such as search and lookups ungated, because a reviewer who approves everything by reflex stops reading the requests.
How it works: langchain 1.4.6 adds per-call interrupts to HumanInTheLoopMiddleware. List gated tools in interrupt_on, set interrupt_mode="per_call", and each gated call raises its own interrupt while a Postgres checkpointer saves the paused run. You resume with one decision per interrupt, keyed by its ID. langgraph-checkpoint-postgres 3.2.0 adds prune, so finished threads can be cut back to their latest checkpoint with strategy="keep_latest". Use reject to deny a tool: a respond message counts as a successful tool result.
Try it: Build an overdue-invoice chaser with two tools, lookup_invoice (ungated) and send_email (gated with True). Run Postgres locally, send each approval request to yourself on Slack or Telegram, try every decision once, including an edit to the subject line, and schedule a nightly prune on finished threads.
TRENDING AI RESEARCH
Harness2: an agent that tunes its own harness for each new task
What it says: Harness 2 (Northwestern University and Google Cloud AI Research, arXiv, 9 October) refines an agent's harness (system prompt, guardrails, memory, skills, tool guidance, subagents) for each new task with model weights frozen. It tests a general and a domain-specific edit against a reference run and keeps only changes the observed behaviour supports. A single refinement step improved every configuration, by up to 11.5 points on WorkBuddy-Bench and 11.8 percentage points on LAB partial pass rate.
Why it matters: The reference run records behaviour without judging the deliverable, so no ground-truth checker is needed. The weaker open model gained most: GLM-5 on OpenCode rose from 64.2 to 75.7 on WorkBuddy-Bench. Scoring leans on LLM judges, and the paper reports cost only as a tokens-per-task chart, with no dollar figures.
How you can use it: Run the loop by hand in Claude Code, Codex, OpenCode or n8n: save a transcript of one common task as your reference, rerun it with a general and a task-specific edit separately, keep edits whose effect shows in the transcript, and log each in a library tagged general or task-specific. Average three runs, as the authors did. Code: google-research/harness2 (Apache 2.0).
TOOL REVIEW
Pancake: outbound prospecting run by your AI agent
What it does: Pancake (#1 Product of the Day on Product Hunt, 14 October) learns who buys from you from your website, then runs Plays: scheduled lead searches on hiring, funding, tech-stack and competitor-engagement signals across 30+ data providers. You approve every lead, messages go from your own account, and your agent drives it through one MCP connection from Claude or Codex. It does not send email outreach, sync to a CRM or make calls.
Price: From $99 a month for 60,000 credits (about 120–600 qualified leads) per workspace, billed monthly, as of 18 October 2026. Plays, messages and replies cost no credits, and there is no overage billing. The 7-day trial gives 15,000 credits and needs a card; payments are non-refundable except where required by law.
Verdict: The first outbound stack built agent-first: worth it if an agent already runs your day, overkill if you send 20 messages a week.
QUICK HITS
DeepSeek released V4.1-Flash-1017 weights under the MIT licence on 17 October (763B parameters, 1M-token context), with API pricing unchanged; benchmark gains are vendor-reported so far.
A Wondery-commissioned study of 3,000 US creator-show fans found 81% trust their favourite host as much as or more than their own friends.
ChatGPT removed the fixed cap on Sites for Plus and Pro on 15 October; usage and storage limits still apply.
That's for today. Hit reply and tell me what you're building. I read every email.
Dev