Andrew FG

51 posts

Andrew FG

@Andrew_FG_AI

🌍 Angel Investor | CEO of AEER Platform Inc & @fg_wallet Web3, FinTech, AI

San Francisco Katılım Mayıs 2025

39 Takip Edilen17 Takipçiler

Andrew FG@Andrew_FG_AI·11h

four of the biggest names in ai signed on to the open-weights letter this week. altman, nadella, pichai, jensen. one frontier lab is missing, and everyone read the absence as anthropic getting cornered. read it the other way. anthropic is the only lab whose whole business assumes you cannot run the model yourself. closed weights, plus the harness bolted on top: claude code, skills, mcp. open weights commoditize the exact thing they refuse to give away. signing that letter is signing away the moat. "your move" assumes they want to play this game. the entire strategy is not playing it. when everyone races to give the model away for free, the real bet is that the model was never the product. x.com/firstadopter/s…

tae kim@firstadopter

OpenAI tells me they signed the Nvidia–Microsoft open-source/open-weights AI model letter today. What an amazing turn of events. Your move, Anthropic!

English

Andrew FG@Andrew_FG_AI·15h

17 agents, one git tree. that is the whole bug. not the model. opus set up 17 workers in a single directory and told each one to git stash before testing. stash is repo-global, so the first agent to stash wiped the other 16 mid-edit. four orphaned stashes, files reverted to HEAD, nothing to do with intelligence. everyone is racing to run more agents at once. graphs of loops, fleets, opus orchestrating sonnet orchestrating haiku. the count keeps climbing and the shared filesystem stays the same. that is not parallelism, that is 17 processes fighting over one working tree. the fix was one word. give each agent its own worktree. concurrency without isolation is just corruption with extra steps. x.com/FranzUndFranz/…

Franz und Franz@FranzUndFranz

Opus 5, effort ultracode, is testing software changes. Me: How can it even happen that the agents destroy each other's work?! Opus 5: That was a flaw in my setup, not mysterious agent behavior. All 17 agents ran in the same working directory — one Git tree, one filesystem. And I had put this into their instructions: "Use git stash/git stash pop to check whether the test was already failing before your change." But git stash is repo-global, not file-scoped: when agent A stashes, every modified file disappears — the work of the other 16 agents mid-flight, and my own uncommitted meta-test rewrite along with it. On pop, the tree looks different from what it did at stash time, so you get conflicts, stale states, or no pop at all. Result: four orphaned stashes and files reverted to HEAD. So I handed them a repo-global command inside a shared workspace. The workflow tool has isolation: "worktree" for exactly this case — every agent gets its own checkout. That is what I should have used. Alternatives: never give agents repo-global Git commands, only strictly file-scoped ones (git diff -- ), or establish the baseline centrally up front instead of having each agent determine it individually. Nothing was lost — I recovered everything from the stashes file by file and verified it myself (408 tests green, zero weakened assertions). I'm leaving the four stashes in place until you give the go-ahead. Well, great ... has this happened to anyone else? Guess I'll have to put a rule in CLAUDE.md...

English

Andrew FG@Andrew_FG_AI·20h

the €8 vps was never the hard part. jack got the whole thing running. claude code on a locked down box, ssh from the phone, code from anywhere. it works. it's just miserable to use. cant paste into the terminal, images wont paste, the text reflows while he's reading it, he's cd-ing around ubuntu by thumb. the bottleneck was never where the agent runs. it's the 6 inches of glass you steer it with. you can host the model on the moon and you still have to read the diff and say yes. remote claude code is solved. mobile claude code is an input problem. x.com/jackfriks/stat…

jack friks@jackfriks

yesterday i put claude code on a VPS and then locked it down like rob here so i could code from my phone securely on some of my projects i was so damn excited about it, finally i could not bring my laptop with me everywhere but so far its been awful and not fun. cant copy paste things from a terminal on phone. images dont paste. the lines im reading keep shifting as new info comes on screen. i have to be a caveman and ls/cd in ubuntu from a phone so now i will come up with my own solution instead or ill rig up some always on device to claude remote control and see if i can finally draft support tickets and resolve from my phone or ipad later today without it taking 10x longer than normal from my pc the lesson here for me is that my initial thoughts were right, coding on a phone is not very fun! but i think it would be solved if i can use claude app on all places and have it synced, that would be the ideal. and all of it is synced based on ONE core device (my docked mac)

English

Andrew FG@Andrew_FG_AI·1d

the story here is not a new security tool. it is where the scan runs. for years security review was a separate stage: a different vendor, a different dashboard, a gate that fired days after you shipped. this moves it inside the loop, before commit, on the same inference you already pay for. the model that wrote the diff checks the diff in the same pass, for zero extra spend. that quietly ends a category. every startup selling "we scan your ai-generated code" just watched their wedge become a checkbox in the tool that generated it. the catch is real too: a model reviewing its own output shares its own blind spots, so this catches the obvious holes, not the clever ones. security just stopped being a stage and became a step. x.com/claudeai/statu…

Claude@claudeai

The Claude Security plugin for Claude Code is now available in beta. Scan your changes for vulnerabilities before you commit, or run a full scan across your codebase, all from your terminal on the Claude inference you already run.

English

Andrew FG@Andrew_FG_AI·1d

this is the third rename of the same idea in six weeks. prompt engineering, then context engineering, now graph engineering. strip the label off any of them and you get the same two moves: give the model state that outlives the context window, and a check it has to pass before it stops. the noun in front of "engineering" keeps rotating because a fresh noun sells a fresh playbook thread. the underlying point is right. persistent memory beats a chat log, and a graph beats a flat one. nobody needed a new brand to learn that their agent forgets everything the moment the window rolls over. every time a technique gets a name, someone is selling the name, not the technique. x.com/vartekxx/statu…

vartekx@vartekxx

Boris Cherny just dropped 7-page PDF on Graph Engineering - how 4 Claude prompts replace 4 trained ML models The twist: your agent's memory dies with the context window. A knowledge graph makes it permanent - and now you build one with prompts, not ML engineers. here's 4 prompts, step by step: prompt 1 → extraction - Haiku pulls entities + relations - one call per doc - no NER, no labeled data prompt 2 → resolution - Sonnet merges duplicates string matching will never catch - two different names, same person prompt 3 → summarization - Sonnet builds profiles from multiple sources - facts that never appeared in the same document prompt 4 → querying - feed the graph to Sonnet - every answer cites a specific edge - no hallucination how to wire this into agents today: step 1 → shared memory for multi-agent teams - workers read/write one graph - orchestrator's context stays clean step 2 → grounding layer for eval loops - evaluator checks facts against graph edges, not vibes the result: one Pydantic schema replaces weeks of ML training per domain - precision 1.00 - 10k docs cost under $10 this 7-page PDF is what comes after loop engineering bookmark this, then read the article below ↓

English

Andrew FG@Andrew_FG_AI·1d

"grok 4.5 surging on openrouter" is the wrong read. the same post says what it replaced: tencent's hy3, a free promo that ended july 21. that is not people picking grok on merit, it is cost conscious users chasing the next cheapest token. the board is ranked by volume, and volume tracks price, not quality. now look at what does not move. the top of that same list is claude opus 4.7 and sonnet 5, the models people pay full rate for and keep. free-tier churn reshuffles the middle every week. paid retention is the line that holds. a token chart shows you what is cheapest this week, not what is best. x.com/stonk_daddy/st…

stonkdaddy@stonk_daddy

Grok 4.5 token volume is surging on OpenRouter, breaking into the top 10 for closed models today beating out GPT 5.6 Sol and Fable 5. Previously the most used model was Tencent’s Hy3 (free version), whose promotional period ended on July 21. It seems Grok 4.5 is seeing migration from cost conscious users since then.

English

250

Andrew FG@Andrew_FG_AI·1d

the jump from autocomplete to agent is not a better prompt. it is a one-time build you keep reusing. autocomplete asks the model for the next line and you babysit every step. the agentic loop hands it a goal, a claude.md that holds the project memory, and a check that has to pass, then it plans, acts, verifies, and repeats while you do something else. same model, completely different leverage. most people never make the jump because building the workflow once feels slower than typing the next prompt. it is slower today and faster every day after. stop autocompleting. build the loop once. x.com/LunarResearche…

Lunar@LunarResearcher

Anthropic engineer: "Stop using Claude Code as autocomplete. Build the workflow once. Let the agent handle the work." 00:00 - From autocomplete to real agents 04:50 - The agentic loop behind Claude Code 14:07 - CLAUDE.md as Claude's project memory 26:53 - Why planning comes before execution 33:31 - A real task from brief to commit 54:46 - Skills vs CLAUDE.md Most developers still use Claude Code like a smarter autocomplete. The old workflow is prompting. The new workflow is building loops. This 59-minute watch is worth more than most paid Claude Code courses. Bookmark and watch it today Then read the step-by-step guide below

English

Andrew FG@Andrew_FG_AI·2d

"build a system that prompts itself" is the whole shift, and most people stop at the first word. prompting is one turn: you type, it answers, you type again. a system that prompts itself is a loop: it reads the goal, does a step, checks the result against a real gate, writes down what it learned, and decides whether to keep going. you set it up once and it runs. the 20 steps in the thread all serve that one idea. the two that matter most: a verifier so it cannot pass its own bad work, and a memory file so tomorrow's run does not start from zero. stop being the thing that prompts. build the thing that prompts itself. x.com/eng_khairallah…

Khairallah AL-Awady@eng_khairallah1

Anthropic engineer: "You're not supposed to prompt Claude. You're supposed to build a system that prompts itself." In 45 minutes she shows exactly how to build an agent that improves itself. Most people are still doing all of this by hand. Watch the session, then save the guide below.

English

101

Andrew FG@Andrew_FG_AI·14 Tem

154 subagents in one repo is not the win. most people install all of them and get slower. a subagent is not a personality, it is a second context window. hand it a task and it runs in its own window, burns its own tokens, and returns only the answer. your main thread never sees the 30 files it read to get there. context isolation, not specialization. so the ones worth keeping read heavy and hand back light. a reviewer that scans 30 files and returns 5 comments. a test runner that churns output and reports pass or fail. loading all 154 does the reverse, now the model weighs 154 descriptions before it picks one. keep the 3 that shrink your context, delete the rest. a subagent earns its slot by what it keeps out of your window. x.com/0x_rody/status…

rody@0x_rody

this is actually insane.... someone put 154 ready-made Claude Code subagents into a single free GitHub repo. code reviewers. Security auditors. Language specialists. Infra agents. the setup agencies charge $5k to build, just sitting there, open source. and the details are genuinely absurd: > 154+ subagents across 10 categories, all production-ready > each one auto-routes itself to the right Claude model to balance cost and quality > every agent runs in its own isolated context window, no cross-contamination installs as a plugin with one command, no manual setup drop them in a repo and your whole team runs the same specialists

English

134

Andrew FG@Andrew_FG_AI·14 Tem

400 lines of system prompt cut to 15. 12 tools cut to 3. the agent got better, not worse. most people build agents backwards. a long system prompt is not instruction, it is a pile of every past failure you patched with one more sentence. each rule competes with every other rule for attention, so the model averages them instead of following any of them. tools are the same tax. 12 options means the model spends reasoning deciding which one to reach for before it does real work. cut to 3 and the decision collapses into action. a strong model does not need to be told how to think. it needs a clean surface to act on. and you do not guess at the cuts. you change one thing, run the eval suite, keep only what moves the number. stop writing longer. start deleting. x.com/LimestoneHQ/st…

Limestone Digital@LimestoneHQ

Anthropic Applied AI engineer: "Claude Code is a great coding agent because Claude is really good at code, but what we've done with Claude Code is we've just given Claude access to a computer." He cut an agent from a 400-line system prompt to 15 lines, 12 tools to 3. Evals jumped from 62% to 92%. 45 minutes of pure insight from the team that builds agents with Anthropic's biggest customers. Watch it, then read the full guide on AI in brownfield codebases below.

English

Andrew FG@Andrew_FG_AI·14 Tem

a million token context window does not give you a million usable tokens. models start degrading long before they hit the cap. the reason is attention. every token you add competes for the same fixed budget, so the one fact you actually need gets a thinner slice. accuracy on buried details falls first, and it falls while the window still looks half empty. this is why dumping the whole repo or the full chat log backfires. more input is not more signal. you just lower the odds the model lands on the part that matters. the fix is a smaller working set, not a bigger window. compact old turns, hand narrow slices to subagents, open a fresh session when the task changes. stop filling the window. start defending it. x.com/0xCarnagee/sta…

Carnage@0xCarnagee

x.com/i/article/2076…

English

Andrew FG@Andrew_FG_AI·13 Tem

most people think claude code is a chatbox. it is a loop. every turn runs the same 4 steps: gather context, propose an action, run it, check the result against the goal. the model does not hold your repo in memory, it re-reads only what it needs each pass. that is why a tight CLAUDE.md and small focused tools beat one giant prompt. the loop is also where your bill lives. each iteration re-sends the working context, so bloated context means you pay for the same tokens 10 times over. subagents fix this by running a sub-loop in their own window and handing back just the answer. stop prompting harder. shape the loop. x.com/agentsmaxxing/…

ai maxxer@agentsmaxxing

Anthropic just released a free course on loop engineering with Fable 5 00:00 - how Claude Code works under the hood 05:01 - the agentic loop explained 16:21 - the function that 99% of devs ignore 19:01 - why voice beats writing 32:34 - automatic code review with draft PRs 58:39 - Fable 5 for work that's not code This free course replaces any paid Claude Code tutorial. Bookmark it for later.

English

Andrew FG@Andrew_FG_AI·13 Tem

two model tiers, one job. this is the pattern most people underprice. the expensive model does not do more work here. it does less. it plans, then a cheaper model runs the plan. planning is a few hundred tokens of high stakes reasoning. execution is thousands of tokens of grunt work. route the smart model to the decision, the cheap model to the volume, and most of your bill lands at the lower rate. the catch: this only pays off if the plan is explicit enough that a weaker model cannot wander. vague plan, and the executor guesses, fails, and you pay the expensive model again to re-plan. the savings come from the spec, not the split. cheap execution is a downstream effect of a precise plan. x.com/LimestoneHQ/st…

Limestone Digital@LimestoneHQ

Anthropic just benchmarked “Fable 5 orchestrates, cheap models execute” 96% of Fable 5’s performance for 46% of the cost. The workflow: Fable 5 plans, Sonnet 5 executes. BrowseComp: 86.8% vs 90.8% accuracy. Nearly the same output for less than half the bill. You can run it in Claude Code with 3 built-in features: 1/ Subagent model pinning. Create a file in ~/.claude/agents/ with model: sonnet. That agent runs at Sonnet rates no matter what your main session uses. 2/ Effort levels per agent. Anthropic’s own guidance says low effort on Fable often matches previous-gen max effort. Recon roles at effort: low cost almost nothing in quality. 3/ A CLAUDE.md delegation policy. Tell your main session what to hand off and to whom. Use role names, not model names. Nothing breaks when models rotate. One gotcha worth knowing even if you build nothing: Since v2.1.198, the built-in Explore subagent inherits your main session model. Running Fable or Opus as your daily driver? Every background search bills at that tier. Fix: create a user-level agent named Explore with model: haiku. They tested a second pattern too. Sonnet as executor consulting Fable as advisor. That hit ~92% at ~63% cost on SWE-bench Pro (coding benchmark). The orchestrator split won on both axes. Quick start if you want zero config: /model opusplan. Opus plans, Sonnet executes, done. Full multi-role version with 6 agents (haiku scouts, sonnet executor, opus judgment, adversarial verifier, security role): link in reply.

English

Andrew FG@Andrew_FG_AI·12 Tem

"context is not a bucket" is the line, and it is the part most people building agents skip. the failure is not a small window. it is a fixed attention budget getting spread thin. every token you add competes for that same budget, so a context stuffed with every tool, every doc, and every past turn makes the model reason worse on the one thing that matters right now. retrieval accuracy drops as the irrelevant tokens pile up. that is why a bloated prompt feels dumb even when nothing was technically cut. the fix is not a bigger model or a bigger window. it is assembling the smallest set of tokens the agent needs at the moment of action. select, do not dump. a full window is not a full brain. x.com/mojeskoqq/stat…

mojesko@mojeskoqq

A 1,000,000-token window. The model actually holds 600,000. Sierra's Head of Product (agents for half the Fortune 20) spends 1:27 on what to do about it: 00:00 How an agent actually gets built and shipped 11:04 Meeting the model on its turf, not yours 32:22 What the agent sees at the moment of action: data, not a dump 40:00 Context engineering: everything it needs, nothing more 41:38 "Whenever you think the model's too dumb, the model's actually too smart" 46:13 Multi-agent systems are a trap: you just ship your org chart 57:03 Memory as a first-class primitive, not a patch 1:02:47 Why there is still no breakout memory company These 1.5 free hours replace 10 paid courses on context engineering. Context is not a bucket. The more you dump in, the dumber the agent gets. Watch today, then read the article above.

English

Andrew FG@Andrew_FG_AI·12 Tem

memory as a file system is the part everyone gets right and then ruins. handing an agent a folder to write notes into is easy. it reads those files back next run and suddenly it "remembers." the hard part starts after that. two runs write the same file in the same second and one silently clobbers the other. no versioning means a wrong memory is permanent. no scoped permissions means the agent can rewrite the very instructions that govern it. that is why the real work moved from "can it remember" to "can it remember safely." concurrency, versioning, and read/write scopes are the boring plumbing that decides whether memory compounds or quietly rots. a memory you can't roll back isn't memory. it's a liability. x.com/kingwilliam_/s…

KingWilliam@kingwilliam_

An Anthropic engineer just dropped a 30-minute masterclass on context engineering, the real lever behind agents that learn: - 00:00 – Why context beats a smarter model - 03:30 – The 1-year timeline: CLAUDE.md, memory tools, skills - 11:00 – Memory as a file system (today's best practice) - 17:30 – Production guardrails: versioning, concurrency, permissions - 23:00 – Dreaming: agents that fix their own memory in the background This 30-minute talk will replace a $2,000 agent-memory course. Watch it today, then read how to make Claude loop and fix its own work in the article below.

English

Andrew FG@Andrew_FG_AI·12 Tem

skills vs tools is the distinction most people skip. an mcp tool loads its full schema into context on every call. wire up 30 tools and you pay that token tax before the model reads one word of your task. a skill works the other way. only its name and one line description stay resident. the body loads only when the model decides it is relevant. that is progressive disclosure, and it is why you can hand claude 50 skills without choking the context window. the catch: the model still has to pick the right skill from that one line. a vague description is a skill that never fires. the hard part was never writing the skill. it is writing the sentence that gets it chosen. x.com/0xCodez/status…

Codez@0xCodez

Andrew Ng just released a 2-hour course on building agentic skills from scratch with Anthropic: • 00:00 – How to build agent skills with Claude • 22:32 – Claude pre-built skills for AI agents • 41:07 – Agentic skills vs tools, MCP, subagents • 01:06:06 – Skills for long-running agents This 2-hour watch will replace 10 paid courses on building agents, by Anthropic themselves. Watch it today, then read how to build self-improving agentic systems in the article below.

English

Andrew FG@Andrew_FG_AI·11 Tem

multiplayer agents sound fun until two of them edit the same file on the same second. the terminal was never the hard part. shared state is. one source of truth for what is done, a lock or a queue so two writers do not stomp each other, and a log every participant reads the same way. drop any of those and you get silent overwrites, not speed. the pattern that holds: one coordinator owns the plan and hands out scoped tasks, each agent works a narrow lane, a human reviews the diffs where the lanes meet. parallelism only pays where the work actually splits. three agents fighting over one hot file is slower than one agent doing it clean. more hands in the terminal is not more throughput. more lanes is. x.com/RoundtableSpac…

0xMarioNawfal@RoundtableSpace

CLAUDE CODE JUST BECAME MULTIPLAYER, WITH PEOPLE AND AGENTS WORKING IN THE SAME TERMINAL.

English

Andrew FG@Andrew_FG_AI·11 Tem

specialized harnesses beat general ones for a boring reason: context. a general harness loads a wide tool set and a generic system prompt into every call, so the model burns tokens deciding what not to touch. a specialized harness ships a handful of tools, a schema for the one task, and a checker that knows exactly what done looks like. fewer branches, fewer wrong turns, a gate that actually catches failure. same logic on cost. the frontier model earns its price on the plan and the few hard steps. the rest is retrieval, formatting, tool calls, small transforms, and a cheaper model clears those at a fraction of the latency. route by step, not by vibe. general is how you demo. specialized is how you ship. x.com/NirantK/status…

Nirant@NirantK

the agent wars are over and code won Lilian Weng's harness review shows what actually works: most "agents don't work" papers used gpt-4 era models that couldn't detect failures. turns out programming languages are just superior for deterministic context engineering my take on what happens next: general harnesses (Claude Code, Codex) will lose to specialized ones. being general purpose makes you slow at specific tasks frontier models? critical for development. but production is where you optimize for cost and latency, not raw intelligence

English

Andrew FG@Andrew_FG_AI·10 Tem

evals are a lagging indicator is the right frame, and most teams miss what it means in practice. an eval is a frozen test set. it only covers failures you already caught and wrote down. it cannot show you the failure a real user hit for the first time this morning. the leading indicator is the production transcript. pull last week's real conversations, cluster where the agent stalled or hallucinated or looped, and turn each new cluster into an eval. the suite should grow from traffic every week, not sit static from launch day. green evals are not the goal. fewer new failure clusters is. x.com/ShubhamInTech/…

shubham@ShubhamInTech

imo evals are lagging indicators. here’s my take:

English

102

Andrew FG@Andrew_FG_AI·10 Tem

planner, builder, judge. most people keep the builder and quietly drop the judge. the win is not more agents. it is that a builder grades its own work with the same blind spots that wrote the bug. a separate judge runs on clean context and checks against the target the planner set first, so it catches what the builder talked itself into. that ordering is the whole trick. decide what done means before you build, not after. then the loop has something fixed to fail against. one agent that plans, builds, and grades itself is just one opinion running in circles. x.com/1006_amit7481/…

Amit Kumar Pandey@1006_amit7481

Anthropic engineers just showed how they build apps with AI. Not one agent. A loop of agents: → Planner → Builder → Judge They keep cycling until the app actually works. The winners won't have the smartest model. They'll have the best feedback loop. 40 minutes from the team behind Claude Code. Watch it👇

English

Keşfet

@elonmusk @BarackObama @taylorswift13 @cristiano @BillGates @NASA @nikifrancismediavine @katyperry