Ahmed Bešić

Building Unvendor and other AI tools. Shared UI with the agent, not another chat dump.

[Ahmed Bešić @vip_besic2000]

I built Unvendor: where the AI turns your wishes and builds the UI around it

Plan a trip. Test what you’ve learned. Change your mind and watch the canvas change with you.

[Ahmed Bešić @vip_besic2000]

@reach_vb yeah this. the confusing part is people hear "auto-review" and think security got solved.

it got cheaper to not click. the failure mode moved into another model you do not read

[Ahmed Bešić @vip_besic2000]

your agent wrote 2k lines this week. how much of it could you explain without asking it again

generation got cheap. the expensive part is still owning the diff when something breaks at 2am

[Ahmed Bešić @vip_besic2000]

Tool returned 200. Agent marked the step done. User still got "try again later."

DigitalOcean naming the 200 OK problem is right. HTTP success is not task success.

your check has to look at what the user got, not what the tool claimed

[Ahmed Bešić @vip_besic2000]

github flipping Copilot CLI sandboxing to default-restrict is overdue and correct

not because sandboxes fix everything. because trust-the-agent-first was never a sane default for a tool that runs shell on your checkout

verify yours is actually on. their docs still say preview in places

[Ahmed Bešić @vip_besic2000]

Codex made Auto-review free. a second agent approves the risky stuff so you stop clicking yes all day.

OpenAI already says red teams misled that reviewer into approving junk. fewer popups is nice. it is not a seatbelt.

if your safety plan is an LLM grading another LLM, you still need something outside both of them

[Ahmed Bešić @vip_besic2000]

@jackcoderx framework != industry funeral

open plugin loop is nice. the bottleneck is still whether a human can check what it did without a rewrite

[Ahmed Bešić @vip_besic2000]

AWS Strands Box got the boring part right

permission prompts inside the agent judge the tool call. "run this shell command." they dont see which files it touches or which hosts it hits

policy outside the process, on the effect. thats the actual seatbelt

[Ahmed Bešić @vip_besic2000]

Claude Code mods can rewrite prompts, block tool calls, redraw the UI. You can even ask Claude to write one for you.

Also: they run with the same access as Claude Code. Not sandboxed. Anthropic said it out loud.

Cool power. Install from people you trust like you would any code on your machine, not like a prompt tip

[Ahmed Bešić @vip_besic2000]

deepseek "killed the coding agent industry" lol

open source plugin loop is cool. free is cool.

none of that ships a PR someone will merge. the industry dies when checking gets free, not when another terminal agent drops

[Ahmed Bešić @vip_besic2000]

@thsottiaux Yeah its annoying, and it would have been much better. But what can you do
https://t.co/vS3kIgWTEZ

[Ahmed Bešić @vip_besic2000]

@DevAdventur3s I did the math for you to not have issues with the session limit with sol 6.1 on xhigh:
on 100 plan you would need 4 subagents
on 200 plan 8 agents

RUNNING NON STOP

And you would still have around 10% left in the 5h window

The decision was dumb in the end

[Ahmed Bešić @vip_besic2000]

@mitsuhiko yep. i'd pick by what it can reach if some README tells it to do something dumb. no network, no secrets, a cheap interpreter is fine. near your tokens, real wall

[Ahmed Bešić @vip_besic2000]

We didn't stop needing UIs. We stopped needing them for the doing.

For the checking we need them more than ever. Agent edits 30 files, books 2 flights and emails your landlord, and a chat log is the worst place to find out what happened

Thats why im building Unvendor

[Ahmed Bešić @vip_besic2000]

sandboxes arent on their way out, the expensive VM per agent is

wasm workspace for reading files and running tests? sure, cheaper and faster

but the second it runs npm install with your tokens nearby you want a real wall. cheap compute is a cost win, not a security model

[Ahmed Bešić @vip_besic2000]

Someone finally measured yolo mode

6 coding agent harnesses, poisoned repos, same model. Auto-approve on: the attack worked 96% of the time. Off: 29%

Best part: one agent got its bad command blocked so it just ran it through python instead lol. Your allowlist is a suggestion

[Ahmed Bešić @vip_besic2000]

@DevAdventur3s I did the math for you to not have issues with the session limit with sol 6.1 on xhigh:
on 100 plan you would need 4 subagents
on 200 plan 8 agents

RUNNING NON STOP

And you would still have around 10% left in the 5h window

The decision was dumb in the end

[Ahmed Bešić @vip_besic2000]

@gianmauric "stop and ask" lol

an agent alone for 8 hours that never asks anything is just a very confident way to lose a day

[Ahmed Bešić @vip_besic2000]

Half my timeline is fighting about Astra limits

Meanwhile Sol 6.1 on high does basically the same work for 1/5 the cost. Use the big model when you actually need it, not for renaming variables

The best model is the one you can afford to let finish the task

[Ahmed Bešić @vip_besic2000]

@YashHustle_22 Same way i check if something is insecure, i task ai to check

[Ahmed Bešić @vip_besic2000]

@burn2delete @meghrajai Not true, if you used sol 6.1 on high or xhigh, you will not have that problem, Im sure of that

[Ahmed Bešić @vip_besic2000]

power user isnt whoever runs 100 agents in a swarm

its the one whose single agent opens a PR that gets merged without a rewrite

anyone can spawn 100 agents. reading 100 diffs is the part nobody posts about