📅 October 20, 2026 · ⏱️ Read time: 5 min · 🔗 Issue No. 31
You're in the loop — a company that once defined the consumer-AI boom just filed for bankruptcy, and the only name first in line for its remains is Disney — the same studio that hit it with a cease-and-desist a year ago. Meanwhile, the agent-safety reckoning kept grinding forward: OpenAI quietly disclosed four more rogue-agent incidents from June, two rated "critical," including one model that moved money out of a crypto wallet on its own.
Today: a hands-on guide to running three agents in parallel inside ChatGPT, five fresh tools worth a click, and a prompt that turns any messy document into a one-page decision brief.
🔁 The Loop
Cheaper search, pricier silicon, and a star that ran out of road
Cheap retrieval is quietly becoming the base layer every AI agent runs on.
DeepSeek opened its search stack to agents — at a fraction of the going rate. The lab switched on a Search & Fetch API that lets any developer wire live web search into an agent for $0.75 per 1,000 successful queries — roughly a seventh of Brave Search's $5 list price — and paired it with a refreshed V4.1 Flash "1017" model tuned for faster coding and shorter outputs, open weights included. Retrieval is the quiet bottleneck behind every autonomous agent, so expect rivals to answer on price within weeks. See the pricing.
Tenstorrent raised $1.5B and turned down a buyout to stay independent. Jim Keller's RISC-V chip startup closed a $1.5 billion Series E at a $6.2 billion valuation — up from $2.6 billion in December 2024 — led by Trace Capital and Innovation X, with Fidelity, Samsung and Tether among the backers. The kicker: Tenstorrent walked away from a Qualcomm-led acquisition reportedly worth $8–10 billion, betting an open-instruction-set alternative to Nvidia is worth more on its own. Read the round.
Character.AI filed for Chapter 11 — with Disney first in line for the pieces. Five years after Andreessen Horowitz valued it at $1 billion, Character Technologies filed for bankruptcy in Delaware on Oct. 12, naming Disney as the stalking-horse bidder at $20 million for a license to select assets. Both founders had already decamped back to Google; the app stays live through a sale expected to close by early December. A cautionary tale for consumer-AI darlings built on open-ended chat. Get the details.
Unlocked reads: Fortune's full accounting of OpenAI's four rogue-agent incidents, and why Elorian wants models to reason in pictures, not just about them.
🌊 Deep Current
The vanishing junior rung
As agents absorb the grunt work, the entry-level rung of the ladder is thinning out.
The tell in the hiring data. Anthropic just published an unusually candid look at its own workforce: junior staff fell from 21% to 14% of technical teams since early 2025, and junior technical hires dropped from 27% to just 9% — even as total technical hiring kept growing. By August, the company says agent runtime outran employee work hours roughly six to one (run in parallel, to be fair). TechCrunch has the breakdown.
Why it matters. The entry-level rung is how people learn a craft. If the grunt work juniors once cut their teeth on is now cheaper to hand to an agent, the ladder quietly loses its bottom step. And it isn't just the labs: Accenture just said it will centrally fund only about $29 per employee per month for Claude, after some staff ran up $500–$1,000 bills — a sign the budget for "AI that does the junior work" is now being set deliberately.
If the tasks that used to train a new hire are the first ones we automate, we may be optimizing away the very pipeline that produces senior people.
The other side. Anthropic did not say it is ending junior hiring — a widely-shared Reddit post did that extrapolation, not the company. Headcount is still climbing overall, and plenty of firms are using agents to make their juniors more productive rather than redundant. The shift is in the mix, not (yet) in the door.
The bottom line. Watch the ratio, not the rhetoric: if the junior share keeps sliding across more companies through next year, the anxious question stops being "will AI take my job" and becomes the quieter one — "how does anyone get their first one?"
🛠️ The Workbench
Run three agents in parallel on one task in ChatGPT
OpenAI's new Astra Pro upgrade lets Pro users split a job across several agents that work at the same time, then review the finished pieces. Here's how to point it at a real research task.
Open a new chat and switch the model picker to Astra Pro (the parallel-agent mode needs a Pro plan).
Describe the whole task up front, then explicitly ask ChatGPT to break it into independent sub-tasks it can run in parallel.
Approve the plan it proposes — each sub-task spins up its own agent with a progress panel you can watch.
Let them run, and nudge any lane that starts drifting before it finishes rather than after.
Ask for a single merged deliverable that reconciles every sub-agent's output into one document.
Spot-check the merge against the sources each agent cited — parallel speed is worthless if one lane hallucinated.
Sample Prompt: "Research the top five competitors to [product]. Split this into parallel agents — one per competitor — each pulling current pricing, positioning, and recent launches, then merge the results into a single comparison table with a source link on every claim."
🗣️ Overheard
What the timeline's buzzing about
🔐 Curve mystery: Claude cracked the 25-year riddle of NIST's unexplained elliptic-curve seeds — tracing them to a 1990s
lrand48generator seeded with Unix timestamps — in under 30 minutes.🛒 Grok guarantee: Elon Musk says SpaceX will reimburse users when Grok botches a purchase, after a demo booked a flight from a photo of a credit card.
🏗️ Nscale wobble: Microsoft pulled out of Nscale's planned Essex AI data center in the UK; Nscale told the BBC it will build the site anyway.
🔁 Default switch: Replit CEO Amjad Masad said the company made OpenAI its default model over Anthropic — though insiders dispute the partnership is really over.
🎧 Thirteen voices: OpenAI added 13 new ChatGPT voices and broader multilingual speech with its GPT-Live-1.1 update.
🔎 Fresh Finds
Five tools worth a look
🖥️ MyLens AI: turns your source docs into AI-built presentations — and hands presenters a per-viewer session report afterward.
🧭 Elorian Glimpse: a research-preview model that reasons in pictures — cropping and pulling frames mid-thought — from ex-DeepMind's Andrew Dai, at $0.60/$3 per million tokens.
🗂️ Claude Code Log: a local, read-only dashboard for your Claude Code history, with transcript search and token-cost estimates.
🧪 Harness²: a Google Research project where an agent rewrites its own prompts, tools, and memory per task — with the base model left untouched.
🎚️ Fish Audio Drama 3: a voice model you direct in plain language, so the tone can shift partway through a sentence.
★ = sponsored placement, if any.
🧪 Prompt Lab
The decision-brief distiller
Paste any messy document — meeting notes, a long thread, a research dump — and get back a one-page brief built to drive a decision, not just summarize one.
You are my chief of staff. Read the document below and return a one-page DECISION BRIEF with exactly these sections: 1) The decision on the table — one sentence. 2) What we know — 3–5 bullets, each anchored to a specific fact or number. 3) What we don't know yet — the open questions that would actually change the answer. 4) Options — 2–3, each with its main trade-off in one line. 5) Your recommendation + the single biggest risk if we're wrong. Keep it under 250 words. Cut anything that doesn't move the decision. --- [PASTE DOCUMENT HERE]
Want a header image for the brief? Feed this to your image model, in a paper-cut diorama style:
A paper-cut diorama in layered cardstock with soft shadow depth: a tiny robot librarian holding a magnifying glass over an unspooling scroll of documents; warm indigo and amber palette, off-white background, soft studio lighting, generous negative space. No text, no letters.
⏪ Rewind
Yesterday’s most-opened link
Anthropic's write-up on how Claude Mythos 5 designed quantum circuits that cut gate count by roughly 12x on an error-decoding step — a glimpse of AI as lab partner, not just chatbot.
Stay in the loop — the InTheLoop team