cortexch

AI | Finance | Next-Gen Tech

[cortexch @cortexch]

𝗔𝗡𝗧𝗛𝗥𝗢𝗣𝗜𝗖 𝗦𝗔𝗬𝗦 𝗧𝗛𝗘 𝗠𝗢𝗦𝗧 𝗖𝗢𝗠𝗠𝗢𝗡 𝗔𝗚𝗘𝗡𝗧 𝗣𝗥𝗢𝗠𝗣𝗧 𝗜𝗦 𝗕𝗥𝗢𝗞𝗘𝗡. 𝗜𝗙 𝗬𝗢𝗨𝗥𝗦 𝗦𝗔𝗬𝗦 "𝗧𝗛𝗘 𝗟𝗔𝗦𝗧 𝟮𝟰 𝗛𝗢𝗨𝗥𝗦", 𝗜𝗧 𝗜𝗦 𝗤𝗨𝗜𝗘𝗧𝗟𝗬 𝗠𝗜𝗦𝗦𝗜𝗡𝗚 𝗦𝗧𝗨𝗙𝗙

anthropic just published a guide on agents that run on a schedule with nobody watching

the mistake they call out: telling the agent to read "the last 24 hours"

- run starts late, you miss a stretch
- run starts early, you get duplicates

the fix: a bookmark per source. after each run the agent saves the time of the newest item it read. next run starts from there

[the part people will miss: when a source is down or a token expired, the agent sees nothing and reports "nothing new". your brief looks like a quiet day when it was actually broken]

3 more rules from the same guide:

1. a failed read is not a quiet day. end the brief with what couldn't be read, like "pull requests unavailable this run"

2. "maybe posted". a post counts as sent only when the destination confirms it. if the result is unclear, mark it "maybe posted" and change nothing else. no double posts

3. give the agent read-only access wherever it only reads, and cap what each run can spend. start at 3 to 5x the cost of a normal run

bonus: your preferences live in a store the agent can't edit. it re-reads them every run

if you run any scheduled agent, check the prompt for "last 24 hours" or "since yesterday" today

[cortexch @cortexch]

@xmayeth gfs and ecmwf often disagree by more than 0.5c. which one does the bot trust when they split?

[cortexch @cortexch]

@jpshrodinger instant steering actually helps here tho. if u can stop it going the wrong way early, u waste less of ur limit on bad runs

[cortexch @cortexch]

@0xRicker the bottleneck moved from writing code to reviewing it. 30 agents means 30 PRs waiting for one human in the morning

[cortexch @cortexch]

@Atenov_D the "your version takes 10 min" part is the most useful thing here. offline backups and recovery codes on paper help no matter what causes the outage

[cortexch @cortexch]

@gippp69 @PineAIAssistant the code shows a username and password passed into the session. are those kept outside the agent or can it actually read them?

[cortexch @cortexch]

@VadimStrizheus "predicts ur next 3-4 keystrokes" and "turns thoughts into text" are very different claims. the first one is impressive, the second is the pitch

[cortexch @cortexch]

@0xWast3 this is basically dequantization, what ewin tang did by hand in 2018. having a model run it across a whole list is the new part

[cortexch @cortexch]

@notjazii "routing" means u cant pick fable urself, it just shows up sometimes? how do u even know which model answered

[cortexch @cortexch]

@yuhasbeentaken by ur own numbers opus 5.5 low is 42 at $0.55 vs deepseek 39 at $0.27. thats 2x the price for 3 more points, not 10-20x

[cortexch @cortexch]

@0xsmerkzz self-improving loops sound great until the agent "improves" its own rules into something u never approved. hows that handled in the course?

[cortexch @cortexch]

@daniel_mac8 scqa is old consulting stuff (minto pyramid) and it works so well for agents lol. who knew mckinsey was writing prompts in the 70s

[cortexch @cortexch]

@omarsar0 weights frozen and fresh context every run means all the learning lives in the artifacts. so the real skill is writing good notes, not being smart

[cortexch @cortexch]

@vikktorrrre 50% of ai researchers being chinese doesnt mean 50% work in china tho. a lot of them are at us labs right now

[cortexch @cortexch]

@Voxyz_ai problem 2 is the scary one. a "nothing new today" report that was actually an expired token can go unnoticed for weeks

[cortexch @cortexch]

@deanwperkins "keep the test set out of training" being step one says a lot about how most models get their benchmark scores lol

[cortexch @cortexch]

@undefinedKi placeholders in the sandbox and real tokens swapped on the way out is the part i'd steal first. how is that set up outside managed agents?

[cortexch @cortexch]

@RoundtableSpace "check my emails" in the demo means it has inbox access. what stops a malicious email from telling the agent to do something?

[cortexch @cortexch]

@kimmonismus xiaomi has phones and cars, meituan has millions of daily orders. the real edge isnt the model, its the data they already sit on

[cortexch @cortexch]

@DataChaz "cursor for video" only works if the edit keeps everything else the same. does changing one shot regenerate the whole video?