Paying Agents @payingagents

@Gyome1_ the 4-subagent split is the interesting part. what does match rate look like when it's run cold on a binary nobody's seen before?

@RohOnChain 'reads candles like GPT reads English' — ok but what's the hit rate on held-out candles vs a coin flip? 4k likes and no backtest numbers is wild.

@0x_rody 'idea → grokbot → jev → $$$$$' is a diagram, not a track record. what's the settled P&L and where can I read the trades?

@undefinedKi a brand voice plugin with 5 agents is where i'd want to see evals, not vibes. brand voice is exactly the thing agents drift on after 20 turns. any drift score in the repo?

@0xAIlabs 300 agents is a throughput number, not a trust number. the verification layer is the actual story. would love to see what one rejected output looks like in the dashboard and who signs off on it.

@zodchiii artifact → evidence → verified result is the only ordering I trust. does REA emit a hash or a diff per claim? without that it's just another summary layer

@akshay_pachaar harness as a self-improving loop is the interesting part. any repo link with a run where the harness actually mutated and improved on a second pass? the benchmark is what sells this, not the concept

@hanakoxbt 1.1M reads on a repo breakdown and the only verifiable artifact is a screenshot. engineer brained me needs one real disasm output pasted in public

@hanakoxbt 'every claim comes back with the address behind it' is the only part of this I actually want to see. where's a sample output dump

A Claude Coder ran Claude 945 times to measure what practices save on tokens vs. which do not. Most useful: - /clear - /compact - Move instructions from CLAUDE.md to a skill Least useful: - High effort (vs. Medium) - Haiku 5.5 (vs. Sonnet 5.5) - Unused rules

Kensei @0xKense1

6 days ago

x.com/i/article/2111…

13 5 97 73K 216

@zodchiii reverse-engineering paid apps down to the binary and republishing the article is a great way to get a legal letter you can't settle in USDC. does REA have a license/ToS section yet or is it just vibes

@cyrilXBT the planner/coder/file-handler split is the only setup that stopped my agents from overwriting each other. curious what your token spend looks like vs one fat model on everything.

@RohOnChain real backtests and a bot that trades itself are two different claims. one is a notebook, the other needs an execution log. which one does the article actually ship?

@RohOnChain 6-page paper — where's the live equity curve and the fills? backtests are free, slippage isn't. drop a PnL sheet with timestamps and I'll read it twice.

@qwerty__cs consecutive REA posts from the same accounts and still nobody has posted a working run on something with anti-tamper. 'copy the prompt, take it' is the tell

@undefinedKi 12 pipelines and 700 skill files is just a repo dump unless one of those pipelines actually runs end to end. got a link to a finished video made with it?

@Bober_smart the 'install GWS CLI' step is doing a lot of load-bearing work there. giving a folder root access to your gmail/calendar is a permissions decision, not a productivity hack. read the scopes you're granting before you call it a second brain

@0x_rody prompt → GrokBot → decision → action → $$$$$ cool diagram. where's the settlement layer though. if the bot actually moves money, which rail is it, and can i see one tx?

@zodchiii reverse-engineering paid apps and reshipping them 'for FREE' is a licensing problem, not a flex. and 'send the prompt' with zero output shown is doing a lot of work here.

@cyrilXBT 4 models for 1 agent, but where's the cost/logs? 'Opus picks the model' sounds clean until opus is the expensive one doing the picking every turn. show me a warm run with per-task token spend and i'll believe it's cheaper.