Synthex

I break down how AI agents actually work. Claude Code setups, prompts, harnesses

[Synthex @Synthexfr]

@monokern hope it saves you a few hours

[Synthex @Synthexfr]

THIS IS A PURE F*CKING GOLDMINE

Claude Code quietly shipped dynamic workflows and most people are still running everything turn by turn

the whole difference is one question: who decides what runs next

subagents → Claude, one turn at a time, every result piles into its context
workflows → a script. Claude writes it once, the runtime runs it, Claude only gets the final answer

that's how you put dozens to hundreds of agents on one task without your main session drowning

try it with one word:

ultracode: audit every route under src/routes/ for missing auth checks

what happens:

→ Claude writes a workflow.js with agent(), pipeline() and parallel()
→ one agent per file, up to 16 running at once
→ edits can run in isolated copies so nothing collides
→ independent agents try to break each finding before it reaches you
→ you only read what survived

the part nobody talks about:

> the run is resumable. stop it, edit the script, relaunch, finished agents return cached results
> press s in /workflows and it becomes your own /command
> fix loops can stop after 2 rounds with no progress instead of burning tokens forever
> /deep-research is already built in if you want to watch one run right now

want it on every task: /effort ultracode, or launch with claude --effort ultracode

heads up: it eats usage fast. test it on one folder before the whole repo

docs: https://t.co/5EIvuTYnwc

save this, you'll need it on your next big refactor

[mirku @mirku21]

https://t.co/PaAzxlyrsH

[Synthex @Synthexfr]

@ScottyBeamIO appreciate it mate. tell me which part you try first, curious what breaks for people

[Synthex @Synthexfr]

@minchoi the brunch one got me. "front camera, one take, no cuts" is doing so much work in that prompt, it's exactly what kills the AI look. saveng these

[Synthex @Synthexfr]

@yurshevv respect for putting the two catches up front, most launch breakdowns skip that part. curious if you've run it on a real repo yet or just compared the numbers

[Synthex @Synthexfr]

@roxitsu love the 1996 domain analogy. i'd add the layer right below templates is just as open: the harness that keeps them reliable. templates get copied fast, a good harness doesn't

[Synthex @Synthexfr]

@0x_kern appreciate it. more workflow breakdowns coming this week

[Synthex @Synthexfr]

@Loofyb0i nine bots and one gate means the slowest agent on the desk is you. next upgrade is deciding what GATE can auto-approve under a dollar limit

[Synthex @Synthexfr]

@0xbobaaa the $1.01 sdk fix is the real flex. newgrounds votibg is the true rest through, players are harsher than any qa. post the score when the 3 days are up

[Synthex @Synthexfr]

@fluid_ai_ $0 per image changes behavior more than budgets. once it's free you stop prompting carefully and start generating 200 to pick one. taste becomes the bottleneck, not cost

[Synthex @Synthexfr]

@0xnineinch "reuse a verified result" is the most underrates line here. half of agent spend is re-deciding things it already decided last month

[Synthex @Synthexfr]

@Klonzu the smartest part of the whole stack is two if statements. hard rules first, models only for whatever survives them

[Synthex @Synthexfr]

@0xGenAi 16 bots for $0 is the easy part. the real cost is the one human who now has to review 16 bots' output every morning

[Synthex @Synthexfr]

$90,000 GOOGLE BILL BECAUSE AN AGENT SHIPPED HIS .ENV

someone stole his keys and ran model swarms on his account. turns out codex had published the .env straight to cloudflare pages

the scary part: it took a different model to even find it

your agent has the same access you do. if you haven't fenced it, you're one deploy away from this

30 seconds of setup in claude code:
> deny Read(./.env) and Read(./.env.*) in your settings.json permissions
> keep .env in .gitignore and out of every build folder
> a PreToolUse hook that blocks any deploy until a secret scan passes
> a spending cap on every API key. today, not after the bill

bookmark this before your next overnight run

[Just_Lingonberry_352 @JustLingonberry]

FINALLY found how my key got leaked and got the $90,000 bill from google

Opus 5.5 said codex had PUBLISHED MY ****ING .ENV TO CLOUDFALRE PAGES ‼️

Meanwhile codex couldn't find anything, refused to explain itself and flagged my chat for cybersecurity risk lol

[Synthex @Synthexfr]

@insomnia_vip no repo, claude came up with that plan on the fly. closest thing you can copy: headless claude -p + xargs fan-out, the session per task. posting the full setup tomorrow

[Synthex @Synthexfr]

CLAUDE CODE JUST HIRED CODEX AS AN INTERN

he asked for one 63-second video. Claude noticed Codex was installed, spun up 161 sessions of it to draw the frames and ran the whole thing itself: RAM, failures, stitching

this is the part everyone keeps missing

the model isn't the product anymore. the orchestrator is

one agent that knows what tools are lying around on your machine beats any single model on any benchmark

[Ananth @Ananth7e]

what the actual fck lmao.

i asked claude code to recreate a 63 second video in a specific animation style, and this is the plan it came up with.

it noticed i had codex installed, and figured out it could use its built-in image generator.

so it launched 161 codex sessions to generate the frames with claude orchestrating the entire thing, managing RAM, handling failures, and stitching everything back together.

this is insane.

[Synthex @Synthexfr]

@0xnineinch he didn't share it, @Ananth7e drop the bill pls, the whole timeline wants to know

[Synthex @Synthexfr]

@TFiirmzee from the screenshot it was 700 frames split across 161 sessions, so ~4 frames per session, around 11fps. the reject rate is the number i'dwant too, that's where the real cost hides

[Synthex @Synthexfr]

@nickcammarata agents reading the neurons of the models that power the agents. the loop is getting very tight

[Synthex @Synthexfr]

@AIatMeta 816B open weights built for agents. the real story is what this does to everyone else's API pricing by next month

[Synthex @Synthexfr]

@skunzu 2026 bingo card didn`t have "claude becomes middle management" either