Most agent swarms burn tokens on two quiet leaks: reading raw junk and recomputing finished work. I turned Iron Giant’s 20-repo thread into an interactive 4-stage cost pipeline with clickable links to every repo. cohear.work/visual-page/grok…
holy f*ck this is smart
a 16-year-old cs student in china worked out how to run 300 grok bots for under $0.02 per bot per week
the same swarm would burn $4,500 a week without it
he cut pointless status checks, stopped rereading full chat histories and killed every endless retry loop
now most bots stay idle until there's real work to do
here are the 20 github repos he used to bring the cost down:
1. markitdown > convert pdfs, office files and html into clean markdown before analysis so the bot reads the content, not the layout
github.com/microsoft/markitd…
2. llmlingua > compress long inputs before the model reads them while keeping the facts, constraints and qualifications the task depends on
github.com/microsoft/llmling…
3. firecrawl > collect web content as markdown or structured data so the workflow keeps only the fields it needs
github.com/firecrawl/firecra…
4. rtk > compress common command outputs before they enter the conversation so repetitive logs don't swallow useful context
github.com/rtk-ai/rtk
5. chonkie > split long documents into chunks so retrieval pulls the relevant passages instead of the whole file
github.com/feyninc/chonkie
6. flagembedding > rerank retrieved passages before generation - strong evidence stays, irrelevant matches leave the prompt
github.com/flagopen/flagembe…
7. mem0 > store and retrieve relevant facts across runs so each task starts with selected memory instead of the entire conversation
github.com/mem0ai/mem0
8. prefect > cache successful task results and reuse them when configured inputs match, skipping reruns of steps that already finished
github.com/prefecthq/prefect
9. gptcache > serve saved answers to similar requests instead of generating again - set matching and freshness rules around the task
github.com/zilliztech/gptcac…
10. graphiti > keep changing facts in a searchable graph and pull relevant context instead of replaying the full event history
github.com/getzep/graphiti
11. portkey gateway > cache matching api responses behind one shared gateway so connected workflows reuse answers across repeated requests
github.com/portkey-ai/gatewa…
12. hishel > add http caching and revalidation to your fetching tools so unchanged responses stay out of repeated download-and-analysis cycles
github.com/karpetrosyan/hish…
13. celery > put background jobs on a worker queue - code handles waiting and dispatch, model calls run only inside tasks that need them
github.com/celery/celery
14. requests-cache > store fetched responses with expiry rules so your scripts reuse valid data before asking the bot to look again
github.com/requests-cache/re…
15. apscheduler > run lightweight checks on a schedule in python and batch nonurgent work before passing useful updates to the agent
github.com/agronholm/apsched…
16. routellm > send simpler requests to a cheaper model and keep the stronger one for work that needs it
github.com/lm-sys/routellm
17. phoenix > trace your ai workflow, inspect repeated calls and compare changes against output quality before keeping an "optimization"
github.com/arize-ai/phoenix
18. pybreaker > pause calls to a service that keeps failing once it hits a set threshold, so one broken dependency can't loop forever
github.com/danielfm/pybreake…
19. semantic router > match requests to predefined workflows with local embeddings instead of paying a large model to pick every next step
github.com/aurelio-labs/sema…
20. litellm > track api spending and set budgets before one runaway workflow eats the whole allowance
github.com/berriai/litellm
connect the tools your workflow needs through the terminal, mcp or your own code - route api calls through the gateway you picked so its controls actually apply
start with the biggest leak: oversized inputs, repeated research, empty checks or failed operations that keep restarting
filter before reading > reuse before generating > check for changes > stop failed loops
then measure finished work against total usage, including the tools running around your bot
the full course below shows how to put the workflow together ↓
1
70
Broke down Gautam’s 14-minute GTA6 clipping tutorial into a 4-stage pipeline, all 10 timestamps, and a quick hardware check (the free PC stack needs an 8GB NVIDIA GPU + Gemini API key). Visual guide here. cohear.work/visual-page/gta6…
Make $2,000–$5,000/Month From GTA6 Without Owning PS5 or Spending $1
Full Tutorial:
00:00 $0 Investment
01:09 What Is Clipping ?
01:52 Tips for Creating Social Media Pages
02:42 Create video In this format
03:09 Finding Content Ideas
04:29 Create Unlimited Clipping Videos for free
09:42 How to Do It on Mobile
10:30 Tips to Grow Your Pages
12:10 How to make money
13:22 How to Automate Everything
65
9,600+ people bookmarked this 102-min video expecting Claude tricks—it’s actually Stanford CS229’s masterclass on how LLMs are built: 6NP compute math, DPO vs. PPO, GPT-4 verbosity bias (22.9%→64.3%), and why 0.2% of FLOPs take 39% of GPU runtime. Full 8-chapter visual breakdown here.
cohear.work/visual-page/clau…
INSTEAD OF WATCHING NETFLIX TONIGHT. Spend 2 hour with this. Claude AI FULL COURSE that teaches you how to BUILD and AUTOMATE anything. The people who watch this tonight will wake up tomorrow with a new skill. Watch it and bookmark it now.
10
I’d automate the check before the trigger. If a bot can’t tell whether the booking exists, scheduling it only makes the errors arrive faster. Here’s the workflow from Lauren’s article.
cohear.work/visual-page/grok…
I think a lot of people aren't using Grok Bot to its full potential, because knowing how to automate things is actually quite a difficult skill to learn. I wrote this post to break down how I approach automating anything with Grok Bot!
22
Baseline tells an agent what it should build. DevTools helps you see what it actually built. I mapped how Chrome’s guidance, runtime checks and public evals fit together.
cohear.work/visual-page/mode…
20
We just launched the all-new CoHear.
One recording → verified summaries, action items, slides, and infographics in 65 seconds.
To celebrate, we’re opening a $100K Launch Reward Pool + free 30-day VIP for new users, and $1 for every friend you invite.
Watch the walkthrough and grab your rewards in the first reply.
🎙Big news: The all-new CoHear is officially live!
⚡Behind every meeting, lecture, and call you take, there’s now an entire AI team working for you — turning raw audio into verified summaries, action items, presentation decks, infographics, and real-time AI answers in 65 seconds.
To celebrate our launch, we’re opening a $100,000 Launch Reward Pool 🎁
✅ FREE 30-Day VIP Pass for new users (limited time)
✅ $1 Cash for every friend you inviteWatch the full walkthrough below & check the first reply to claim your rewards in 60 seconds 👇
1
34
Three meeting-note apps all summarize your team sync. Almost none preserve what you rejected.
A summary says "pricing approaches were discussed." A decision record says the discount was killed — and why.
Compared them:
bakucohear.substack.com/p/fe…
25
Asked ChatGPT for the best note app for team syncs with decision summaries.
It named five tools and a comparison table — then admitted: "qualitative assessments based on published feature comparisons, not a controlled accuracy test."
A comparison table is a claim, not evidence.
56
Your team sync produced a summary, not a decision.
A summary says "pricing tiers were discussed."
A decision ledger says: shipped / rejected (and why) / still open, with an owner and the transcript line.
Five-minute test in the post:
bakucohear.substack.com/p/yo…
1
21
An AI summary turns "we should look into pricing" into "team will review pricing."
One is a suggestion. The other is a commitment, with an owner and a date.
Test: can you point at the transcript line each bullet came from?
bakucohear.substack.com/p/a-…
12
A lecture recording gives you a transcript, not practice questions you can trust. Three passes: list the concepts, write recall / explain / apply questions, then verify every answer against the lecture. The answer key is where it usually goes wrong. bakucohear.substack.com/p/le…
8
Before you record a meeting, say three things out loud: that you are recording, why, and what happens to the file. Then say the date, the topic, and who is in the room. Ten seconds of context saves an hour of guessing later.
8
A lecture transcript tells you what was said. A useful study session asks what you can recall, explain, and apply. Keep the source close enough to check the answer.
9
Turning meeting audio into slides is not the same as recovering slides from a screen recording. One creates a new recap. The other retrieves an existing visual. Choose the workflow before the tool.
15
An AI meeting summary should not turn “we could” into “we agreed.” Separate suggestions, decisions, and commitments before building the action list.
13
I just published AI Meeting Notes: Which Workflow Fits In-Person Meetings, Online Calls, and Existing Recordings? medium.com/@basicxyz1029/ai-…
16
Want 1 month of CoHear Pro for free ($14.99 value)? 🎁
It’s an AI voice recorder that turns meeting & lecture recordings into clear summaries and action items.
Reply to this tweet or DM me, and I’ll send you an iOS promo code! 👇
The meeting ended. What did we actually decide?
A quick summary is useful. A clear record of decisions and next steps is even better.
See how CoHear turns a conversation into notes your team can act on.
Download CoHear: apps.apple.com/us/app/cohear…
12
Join via public link is the constraint worth noticing. The meetings that actually matter often have no link to join - the room, the car, the client's office. An agent that has to be let in can't record those, and that's most of the real ones.
A suggestion for the @Grok Grok Bot team. I'd love my bot to be able to join a Zoom or Teams call with me, or for me. Next month I'm listening in on the FAA's public meetings on Starbase launch cadence. I won't be asking questions. What I want is a read on the attitude and tenor of the call, and my bot would be perfect for that. Even just listening and reporting back afterward would be dynamite. Joining as a real participant would be the next step.
25










