AI & tech researcher | Data engineer | Building cool stuff | Sharing everything I learn
Chicago, IL
- Tweets 13,160
- Following 514
- Followers 20,017
- Likes 21,890
Your coding agent is quietly burning money on tools it never touches.
Microsoft just released a plugin that fixes it, it installs in about 2 minutes.
It learns from every session in your repo and makes the next one cheaper. Up to 30.9% lower cost on SWE-Bench, with comparable pass rate. Works with Claude Code, Codex, Copilot CLI and Pi.
github.com/microsoft/TokenEc…
53
5
79
6,187
someone under my 10 ways post asked which one is easiest to land clients for
review replies
how i'd get the first one:
- bot finds local shops with unanswered google reviews
- drafts 3 replies to their worst one
- i send it as the pitch
bot qualifies, human sends
also a fix: one stream is ~$250/week, not $1000
you stack 3-4
10 ways to make $1000/week with grok bot (realistic):
1. local business review replies
a bot drafts replies to every google review for restaurants and clinics. 10 clients at $25/week each
2. real estate listing writer
agents send photos and specs, the bot writes the listing and social posts. charge per listing
3. ecommerce product descriptions
a bot rewrites 100 product pages for shopify stores. sell it as a one-time package
4. price and stock monitor
a bot watches competitor prices and pings store owners when something drops. monthly subscription
5. job application bot
it finds fresh openings and tailors a resume and cover letter for each. job seekers pay weekly
6. meeting notes for agencies
the bot joins calls, writes summaries and action items. sell it per seat
7. newsletter curation
a bot finds the week's best news in a niche and drafts the issue. charge creators per issue
8. invoice and receipt sorting
it reads email receipts and sorts them for freelancers before tax season
9. podcast clip finder
the bot pulls the best quotes and timestamps from every episode. podcasters pay per episode
10. inbox triage for founders
it sorts the inbox, drafts replies and flags what's urgent. busy founders pay for this fast
this is my version of the list, i just stealing the idea and suggested my own ways
19
4
53
2,063
This is the most detailed guide to AI agents on the internet.
If you want to really understand agents, how they're built, how they remember, plan and act, this is where everything is in one place.
A 42-page survey by Renmin University, updated through 2025. It reviews hundreds of agent papers and boils them down to one framework.
Once you see that, every new agent framework stops looking new. It's the same four parts, just built differently.
Key pages:
> p.4 Fig. 2: the whole agent architecture on one picture
> p.5 Memory: how agents remember, read, write and reflect
> p.10 Planning: single-path vs multi-path reasoning, with and without feedback
> p.13 Action: tools, goals and how actions change the environment
> p.16 Capability acquisition: fine-tuning vs prompting vs mechanism design
> p.22 Fig. 5: where agents are used and how they're evaluated
> p.31 Challenges: what still breaks
27
14
143
8,992
This is GOLD
Anthropic just published the code review setup the Claude Code team uses.
Merged PRs more than doubled, reverts dropped by over 40%, and PRs with real review comments went from 16% to 54%. Here is how it works.
1. Review starts before the PR
Claude runs the tests, build and linter and shows the output instead of saying "done". Then a second agent reviews the diff in a fresh context.
2. A REVIEW.md file
Written by code owners. It says what failure looks like in that part of the repo, caps the nits, lists paths to skip and asks for file + line as evidence.
3. Agents vote on bugs
Several agents hunt for bugs in parallel, a separate group votes on each finding. It only gets posted if enough agree it's real. Style nits get dropped.
4. Let the agent review the agent
If you're the only human reviewer, a fresh-context agent with REVIEW.md catches what you skim past.
5. Every finding must be answered
Fix it or close it with a reason. The PR can't merge while one is open.
How to copy it:
- Short CLAUDE.md with the checks your agent can run
- Add REVIEW.md and move repeated review comments into it
- Run /code-review and /security-review before every PR
- Add "Testing", "Risk" and "How to revert" to your PR template
- Cap review rounds, after a few re-review only the new diff
36
37
450
68,931
source: claude.dev/blog/solving-code…
1
10
1,132
You're overpaying for Claude Code. Here's how I cut the cost by ~40% without losing quality.
In one session I ask for a refactor, then a file rename, then a quick explanation. All of it runs on the same expensive model.
Then I came across jev-router and tested it myself. Before each message, a small model looks at it and picks the cheapest one that can handle it: Haiku, Sonnet, Opus or Fable.
I ran it on fastapi, same bug fix plus a regression test:
Opus 5 only: $0.421
Sonnet 5 only: $0.205
jev-router: $0.254.
The router itself cost $0.0003 for the whole test.
All three passed. 40% cheaper than Opus, same result.
Setup:
1. npm install -g jev-router
2. Put JEV_API_KEY from TypeSafe in ~/.jev-router.env
3. Run jev-claude instead of claude
4. JEV_ALLOW_FABLE=1 to add Fable
github.com/gargpratyush/jev-…
22
4
62
4,066
Agent harness, explained like you're five. Every big lab is shipping one, almost nobody says why you need it
So, imagine a horse, it's your model. Strong, fast, and it runs wherever it wants.
The harness is the reins, the saddle and the cart. Without it, all that power goes nowhere useful.
Claude in your browser and Claude Code run the same model. But one answers questions, and the other actually builds a whole projects on your computer. The difference is the harness.
In Claude Code, that's the CLAUDE.md file with your rules, the memory of your project, access to your files and terminal, tools and a loop where it runs the code, checks the result and fixes it.
Here is why you need one. Without it, you are the harness. You repeat the rules, paste the files, check every answer and decide the next step. Every day.
Google researchers kept the model and only changed its harness. On one real office task, the score jumped from 43 to 78 out of 100, just from adding a rule to double-check its own work.
Three things make a good one:
1. Rules written once. A CLAUDE.md or AGENTS.md file it reads before every task, so you stop repeating "use our template"
2. Only the tools the job needs. Every extra app you connect is another way for it to get confused
3. A check before "done". Tests for code, a checklist for everything else. If it can't verify the work, you will have to
Start small. Take the instruction you paste most often and save it as a rule. Take the mistake it makes most often and turn it into a check.
The model is not the bottleneck anymore, it got commoditized. The harness is where the results come from.
39
56
474
30,568
This is genius.
I cut my agent's cost per code review by 49% with one simple setup change.
Cloudflare published how they took their AI code reviewer from $2.43 to $0.77 per run.
77% of their bill was the cache. The shared prompt never got cached because per-job details sat in the same block.
The fix: static part first, dynamic details after, and cut wasted calls.
I tested it on 5 real FastAPI pull requests with Claude Code:
> Moved the rules into a stable prompt. Branch and date go after the diff.
> Cut tools from 16 to 4 and dropped unused MCP connectors.
> No final summary.
$0.144 → $0.073 per review. Cheaper on every single PR.
Both setups found the same core issues.
Wrapped it all into a skill so you don't have to set it up by hand, just one command: npx skills add undefined-ui/cache-tune -g
35
17
219
16,296
My skill repo: github.com/undefined-ui/cach…
Cloudflare's blog: blog.cloudflare.com/obsessiv…
1
10
712
Uber shared the exact agent setup behind 70% of their pull requests.
Weekly users grew 7x, and cost per session still fell 52%. Here's what you can copy.
The setup: coding agents that run locally and in the cloud, used by engineers and non-engineers.
They have 3,600+ shared skills and 30K+ skill runs a day.
The main model plans and checks. Cheaper subagents do the heavy work.
The savings came from defaults, not from a smarter model.
What to copy:
1. Compact early Auto compaction at 400K tokens, even on 1M context models.
2. Lower the default effort Reasoning effort is set to Medium. Turn it up only when a task needs it.
3. Cheaper subagents The main model plans. Subagents on a cheaper model do the reading and searching.
4. Longer cache for the main session Prompt cache lasts 1 hour for the main session, 5 minutes for subagents.
5. Don't load every MCP tool With 100+ tools installed, the schemas alone cost 50K to 70K tokens before the first prompt. Call tools through a CLI and search for them only when needed.
6. Batch tool calls into one script That cut tokens by 55% to 71%, even on small queries.
Start with one default today: compaction, effort level or subagent model.
41
41
344
22,933
source: uber.com/us/en/blog/efficien…
7
696
Your agent gets dumber the longer it works on a task. Just one simple habit fixes this.
Researchers tested three ways to give agents more compute: run one agent longer, run several in parallel and let parallel agents share ideas.
All three hit a ceiling.
The agent gets anchored to its own past attempts. The more history it sees, the smaller its edits get.
The fix is to make it forget. Every so often, close the session, keep only the best result so far, and start a fresh session from it.
On the same budget, that beat a single agent by up to 39% and parallel agents by up to 15%.
To do it yourself: when your agent starts going in circles, ask it to save the best version and what it learned, then open a fresh session and give it only that.
30
5
127
6,411
Paste this prompt when your agent starts going in circles. Then open a fresh session and give it only the handoff:
---
<task>
You've been working on this for a while. Before we restart in a fresh session, write a handoff note.
</task>
<rules>
- Keep only the single best version so far. Paste it in full.
- Do not describe failed attempts step by step.
- No history, no apologies, no summary of the chat.
</rules>
<handoff>
1. Goal: one sentence on what we're trying to achieve.
2. Best version: the full current best result.
3. What works: up to 3 short points on why this version is the best.
4. Dead ends: up to 3 approaches that didn't work, one line each, so the next session doesn't repeat them.
5. Next step: the one most promising thing to try next.
</handoff>
7
349
Hire Viktor as your team's new AI employee. Free to start: ref.viktor.com/yarchi-x-5
8
1,313