In partnership with
Winning Startups Aren't Bigger. They're Leaner.
Founders deploying AI agents across sales, marketing, and customer service are closing more with smaller teams. 65% more sales leads. Less headcount.
Download the free Practical Guide to Agentic GTM for Startups and start building the stack your competitors dream of.
Welcome
This week's announcements point to a practical shift. Grok Bot and Hermes Bot Mode turn one assistant into named specialists that can keep recurring work moving. Google and OpenAI are also making the wait between a question and a usable next step much shorter.
The useful test is the same in each case: start with one defined job, keep the permissions narrow, and measure repair time as well as model spend.
Quick Hits
Five things worth trying this week.
1. Grok 4.6 is live across several developer surfaces. xAI says the model is available in Cursor, Grok Build, its API, OpenRouter, Vercel, and Cloudflare. Test one long-running coding or research job against the model already in use. Read more
2. Meta released Muse Glimmer as an open-weight local model. The 30B model is Apache 2.0 and designed to run on a recent Mac or a single consumer GPU. Use it for a private local experiment before sending the job to a paid cloud model. Read more
3. Gemini 3.7 Flash is Google's new high-volume workhorse. It is available in the API and AI Studio with introductory pricing through year-end. Benchmark one repeated task before changing a production route. Read more
4. ChatGPT desktop activity memory and ChatGPT for Teens are rolling out. Activity memory is opt-in, while the teen experience is age-gated. Review privacy settings and account eligibility before relying on either feature. Read more
5. GPT-5.6 Sol Ultrafast is a select-customer API preview. OpenAI says it can reach up to 14 times standard speed. Reserve it for work where a person is actively waiting on the result. Read more
Top Updates
1. Grok Bot makes a recurring job into a named AI teammate
Grok Bot gives each named agent a persistent cloud computer with a browser, filesystem, and terminal. The intended use is a specialist that can keep moving on a recurring task after the laptop is closed, while asking for approval when a decision or permission boundary appears.
Start with one named job with clear inputs and a review point: lead research, inbox preparation, invoice follow-up, or a weekly research brief.
Why it matters: I can assign a recurring job to a persistent assistant, then judge it by the amount of attention it returns rather than the number of tabs it opens.
Action items:
Choose one repetitive, reversible job with a visible completion test.
Give the Bot only the tools and sources it needs for that job.
Measure interventions, review time, and repair time for three days.
2. Hermes Bot Mode turns existing agent profiles into a reusable roster
Hermes Bot Mode turns existing Hermes profiles into named Bots with their own role, model, memory, skills, and profile picture. Bots can exchange work through a persistent Agent Inbox and hand tasks to each other with mentions.
The benefit is reuse. A team can build a researcher, writer, or operations profile once, then return to the same specialist instead of recreating its prompt and context.
Why it matters: I can keep specialist roles and context separate, which makes a small personal AI team easier to review and improve over time.
Action items:
Convert two existing profiles into Bots with narrow, distinct roles.
Give them a small shared task with a human reviewer and a clear stopping point.
Keep credentials and sensitive data out of shared profile exports.
3. Gemini Flash and Sol Ultrafast make latency and cost part of the workflow design
Google made Gemini 3.7 Flash generally available on August 13 for coding, web, and agent workflows. OpenAI previewed GPT-5.6 Sol Ultrafast the same day for select API customers, promising up to 14 times standard speed and up to 750 output tokens per second.
Flash is aimed at high-volume work where a modest response that arrives quickly can keep a workflow moving. Sol Ultrafast raises the bar for interactive tasks where the team is waiting on the model: a coding loop, incident response, or source-heavy research pass.
Why it matters: I would select a model only after timing the same work, counting rejected output and repair time, and calculating the total cost of an accepted result.
Action items:
Run the same 20-task sample through your current model and one faster candidate.
Record median latency, the rate of accepted outputs, repair minutes, and cost per accepted task.
Route routine high-volume work to the lowest-total-cost model, while reserving limited preview capacity for genuinely latency-sensitive loops.
Pro Tip
Turn a messy decision memo into a private ChatGPT Site
In ten minutes, turn a non-sensitive launch or vendor decision into one readable internal page: decision, options, source links, owner, and deadline. Ask ChatGPT Sites to turn the memo into a simple page, then review every claim before sharing it. Keep customer data, card data, health information, and private contracts out of the source material.
Action items:
Start with a decision that can be shared internally without sensitive data.
Require a section for assumptions, source links, and the person who makes the call.
Set an expiry date and unshare the page after the decision.
Productivity Gem
Use ChatGPT Library to make a launch folder answerable
Connect the Google Drive folder for a launch, renewal, or account plan. Ask for a one-page brief that names the documents used, conflicts between them, and missing decisions. Open the cited source files before using the summary. This is most useful when the bottleneck is finding the current version, not writing another document.
Action items:
Connect one time-boxed folder, not the entire company drive.
Ask for source links and contradictions, not a generic summary.
Assign one owner to resolve each missing or conflicting item.
One idea shouldn't take six rewrites to post.
Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.
Health Tip
Turn one week of activity evidence into better clinician questions
For eligible U.S. adults, collect a week of ordinary activity notes or read-only Apple Health trends in Health in ChatGPT, then ask for a short list of observations and questions to discuss with a clinician. Pair that with one environmental change from NIH's current movement guidance: a walking meeting, a visible pair of shoes, or a regular route that makes movement easier.
Do not ask AI to diagnose symptoms, choose treatment, or handle urgent concerns. Use it to organize facts and questions, then take those to a qualified professional.
This information is for educational purposes only and is not medical advice. Please consult a qualified healthcare professional before making changes to your health routine.
Kids Tip
Build a paper model and test its predictions
Day of AI launched a free PreK-12 AI literacy curriculum and family toolkit on August 19. Pick ten safe household objects, let the child sort them by a visible feature, and have them write a rule that predicts where a new object belongs. Keep the wrong prediction. That is where the learning starts.
For ages 8 to 10, use two obvious groups. Older children can add a borderline object and discuss how examples can produce an unfair or unreliable rule. An adult should handle any optional AI prompt, keep names and photos out of the chat, and ask the child to explain the model in their own words.
This week's test
The fastest model is not automatically the best model. The cheapest model is not automatically the lowest-cost model. The real test is whether it gets you to an accepted result with less waiting, less repair, and less staff attention.
This week, choose one workflow that happens often enough to measure. Test it with the model you already use and one faster option. Count the cleanup. Then make a decision based on the work, not the demo.
PromptHacker Archive
Browse practical prompts, decision checklists, and family-safe AI activities from the archive.