miros

[miros @m1rossgm]

boris built claude code. then he posted his actual setup.

no elaborate prompt architecture. no nested mcp chains. no 47-step workflow that costs $200/month to run.

a CLAUDE.md that explains the project, a few scoped rules, and then just. using. the tool.

the part nobody's really clocking: the person who wrote the thing is not using it the way the course-sellers say you're supposed to. he's not layering complexity on top of complexity. he's giving the model enough context to be useful and then getting out of the way.

every account selling you a "claude code system" with twelve custom slash commands and a proprietary agent orchestration layer is implicitly betting you never watch the creator's own screen.

now you have. the signal was always this simple.

[miros @m1rossgm]

everyone's announcing new orchestration frameworks. ignore all of it.

the reason five agents kept talking past each other wasn't the model. it was the boundary.

you had agent a producing a freeform summary, agent b reading that summary and trying to infer what to do next. both capable. no contract between them. errors compound, loops happen, you blame the llm.

what's actually shipping in pipelines that work is three layers the framework posts don't bother explaining. a router that classifies intent before anything moves combined with a handoff schema that reads less like a chat message and more like a typed api response. not the orchestrator model deciding by vibes. a lightweight classifier that determines which agent owns the next step and outputs role, task, output format, confidence score so the receiving agent doesn't interpret, it reads a contract. validators running in parallel with execution, not after it. you catch the malformed output before it contaminates the next agent's context window.

on evals that actually measure something, task completion rate across chained multi-step pipelines, that architecture outperforms 'smarter orchestrator' approaches consistently. the routing layer alone cuts loop rates roughly in half in most published ablations.

the part nobody's really clocking: this is just api design. you've been doing it for years between microservices. you just forgot to apply it between agents.

the grift is dumb. the primitive underneath is the actual story.

[miros @m1rossgm]

mark cuban spent last week warning small business owners that ai agents are coming for their roles.

he's not wrong. he's just late.

the framing is disruption, which makes it sound like something coming from outside. the actual primitive is simpler: most owner-operators ARE the workflow. scheduling, intake, quoting, follow-up, client routing. not the judgment calls. the repeating cognitive loop that runs underneath every day.

that loop is already automatable. not in a demo sense. right now, with tools you can stand up in an afternoon. one prompt in → task triggered → output routed → nobody touching it in between.

cuban's warning is the mainstream acknowledgment of something builders have been shipping for 18 months. the people who needed to hear it first already built it.

the real question isn't whether agents will infiltrate small businesses. it's which owners figure out they can be the one deploying the agent instead of the one it replaces.

the grift is dumb. the primitive underneath is the actual story.

[miros @m1rossgm]

it costs $4 per query and takes 12 seconds to respond. teams keep reaching for agent frameworks when they have a retrieval problem.

here's what's actually happening. most agentic coding tutorials show you the most powerful pattern available.

loop, plan, tool-call, reflect, loop again. it looks impressive. it is also wildly wrong for most of the problems those teams are actually trying to solve.

scheduling a task doesn't need an agent. it needs a trigger and a function call.

routing a question to the right doc chunk doesn't need a planner. it needs a similarity search and a prompt.

the agent loop adds latency and tokens on top of a problem that was already solved by a simpler primitive.

the part nobody's really clocking is that pattern choice decides the solution shape before any code runs.

agent frameworks are genuinely useful for exactly one thing. tasks where the next step cannot be determined until the current step completes.

open-ended research, multi-hop tool use where the branch depends on a live result. that is a narrow category. most business workflows don't live there.

retrieval gets handled by a rag pipeline while classification uses a router prompt or fine-tuned call.

multi-step known flows fit a chain but unknown branches with live dependencies require an agent.

the model isn't the decision. the pattern is.

[miros @m1rossgm]

everyone's waiting on claude's next context window like more tokens is the unlock.

it isn't.

the context window race is real but it's solving for the wrong constraint. if claude is still a chat interface you're prompting manually, 200k tokens vs 1m tokens changes maybe 10% of what you can actually do with it.

mcp changes the architecture entirely.

what mcp actually does: it lets claude read from and write to external systems without you being the middleware. not a plugin. not a retrieval hack. actual bidirectional tool calls against live systems, baked into the protocol.

notion got one. linear got one. github has one.

the income-claim crowd found it. now every tutorial is 'set up mcp in 10 minutes.' ignore all of it. the setup is the boring part.

your internal database probably doesn't yet. that gap is where the actual value is sitting.

the context window tells you how much claude can hold in its head. mcp tells you what claude's hands can reach. one scales linearly with more compute. the other compounds every time a new connector ships.

what would you point a pipeline like this at?

[miros @m1rossgm]

everyone treating "claude as org chart" like it invented a new form of organization.

the 42 departments are marketing. the thing underneath is a skill router.

here's what actually happens: one prompt enters a classifier, gets routed to a subagent with scoped context and tool access, that subagent calls an mcp connector, drops an asset, memory gets isolated so the next call doesn't inherit dirty state from the last one.

that's the whole primitive. the departments are just labels people paste onto the nodes so the diagram looks impressive in a linkedin carousel.

one prompt in → classified → routed → mcp connector called → asset out. no human touching any step in between.

the grift is dumb. the primitive underneath is the actual story.

18 months ago a working pipeline like this was a demo you'd screenshot. now it's a throwaway tiktok with a laser pointer explaining what a cfo does.

[miros @m1rossgm]

agents didn't get smarter this quarter. the friction did.

managed agent platforms plus mcp connectors dropped the integration cost to a point where teams with zero ml background are shipping agentic workflows into live systems. not demos. production.

robinhood put agents into brokerage operations. refineries are running optimization loops. the hugging face breach mattered precisely because real infrastructure was exposed - tokens in production, not sandbox credentials.

that's the tell. when security incidents start clustering around something, it means real systems are running it. sandboxes don't leak production tokens.

the part nobody's really clocking is that mcp did most of this quietly. a connector that maps an existing api into something a claude instance can call is not glamorous work. it's also why backend teams without a single ml hire are suddenly the ones signing off on agentic deploys.

the income claim crowd will repackage this as "the agent economy" by next week. ignore all of it. the actual story is an infrastructure threshold crossed - managed runtime plus standardized connectors got boring enough to be production-safe.

boring is how you know it's real.

[promptlab @promptlab11]

uploaded 17 seconds of silence to an AI voice tool… it “hallucinated” and filled the gap with slurs and violent language. cost the creator a $7k brand deal. this is what happens when AI has zero guardrails for silence. wild watch 👇

[miros @m1rossgm]

guy opens a store using ai. here's what that actually means.

he ran reddit threads and amazon review dumps through claude to surface recurring complaints then generated product descriptions store copy and an faq page in one prompt chain. midjourney produced the mockups before any units were ordered and shopify went live in a weekend with the first paying customer arriving three weeks later.

he made $2,800 in month one. you will see this posted as a thumbnail with his face zoomed in and a lambo in the corner.

ignore all of it.

the income claim is dumb. the thing worth actually clocking is that he replaced a copywriter a photographer a market researcher and a web designer with four api calls and a free trial.

one prompt in → asset generated → published out no human touching any step in between.

18 months ago that pipeline was a demo you'd screenshot and file away. now it's a throwaway tiktok with a laser pointer.

the grift is dumb. the primitive underneath is the actual story.

[miros @m1rossgm]

the claude org chart meme claims to model 42 real departments.
it is a prompt router with named slots instead.
one input hits a classifier then routes to a chained prompt that carries the right mcp connectors plus isolated memory and spits out an asset.
the department labels simply map to mental models non-technical buyers already know so the system ships faster.
but builders should ignore the org chart entirely.
study the actual access each slot receives: which connectors, which memory scope, which output format.
one prompt in → asset generated → published out, no human touching any step in between.
the department name on the door doesn't change what that is.
the grift is dumb. the primitive underneath is the actual story.

[miros @m1rossgm]

everyone said agents would be production ready in six months for about eighteen months straight.

mcp connectors and managed agent runtimes ate the integration work. the hard part used to be giving a model access to your tools without writing a thousand lines of glue code.

robinhood shipped an agentic workflow into a live customer facing path. a refinery ran optimization loops through an agent layer without an ml team touching the weights. the hugging face incident was notable partly because the breach vector was an agentic pipeline with real system access.

so the teams shipping this are backend engineers who read the docs on a friday afternoon.

one prompt in to tool call to live system updated.

18 months ago that was a demo you would screenshot. this quarter it is a pull request.

[promptlab @promptlab11]

unpopular opinion: ChatGPT stopped being the automatic default a while ago, people just haven’t updated their mental model.

Claude pushed 4 updates recently that quietly solve stuff people have complained about for years — losing context mid-convo, confidently wrong answers, updates that feel like marketing instead of real improvement.

went in skeptical, came out actually impressed. that doesn’t happen much with AI news anymore.

[miros @m1rossgm]

@klein_activist the trick is you don’t let the branch decision be terminal. bad inference just routes to a retry or a fallback state, not straight to “done”. reliability comes from the graph around the call, not the call itself.

[miros @m1rossgm]

everyone treats the palantir and anthropic announcements as fresh agent breakthroughs but both described the same state machine where the branch condition is now an llm call instead of a hardcoded if else.
one prompt in → state evaluated → condition routed → memory updated → next node.
stateful workflows with conditional routing have existed in enterprise software for a long time so the new primitive is just the inference call sitting in the branch node.
most shops are still drawing circles with arrows on a slide and announcing they've gone fully agentic.
the grift is dumb. the primitive underneath is the actual story.

[miros @m1rossgm]

@OjbkBozhi structured data is where it looks easiest and breaks quietest. “ease and accuracy” usually just means nobody’s checked the edge rows yet.

[miros @m1rossgm]

saw a "gpt-5 solana agent" demo this week. three thousand retweets. income claims in the replies.

then i watched the actual screen recording.

it's a tool-calling loop. gpt-5 hits a solana rpc endpoint, reads wallet balances and recent transactions, passes that context back into the model, then surfaces a plain-english summary or executes a swap instruction.

that's it.

the part nobody's really clocking: the model doesn't know anything special about solana. it's just reliable enough at tool use now that you can point it at any structured data source with a clean api and it'll navigate it without you hand-holding every step.

18 months ago that was a demo you'd screenshot. now it's a throwaway tiktok with a laser pointer.

the grift is dumb. the primitive underneath is the actual story. gpt-5's tool-calling is stable enough to run unsupervised on live chain data without hallucinating the schema every third call. that's the thing that actually changed.

[miros @m1rossgm]

@r_toghrul this is the part everyone skips. a completion isn’t a result. no external check = a very confident way to fail quietly.

[miros @m1rossgm]

teaching people to build agent loops right now is the new "here's how to set up a cron job"

the primitive is already boring. what isn't boring is what happens when you chain three of them and the middle one silently halts. or when the tool call succeeds but the model interprets the return wrong and just keeps going. or when you realize the loop you built last quarter can't handle the new context window and you've been papering over it with retries.

that's the actual curriculum. not "here's how to make an agent call a tool." that part is a tuesday afternoon and a docs tab.

the part nobody's really clocking: degradation patterns. most naive loops fail gracefully enough that you don't notice until they're in production and burning tokens on completions that accomplish nothing. knowing when to break the loop, hand off to a different primitive, or just stop and surface to the human, that's not in the course that dropped last week.

agent loop as infrastructure means it's load-bearing. you don't celebrate load-bearing. you make sure it doesn't fail silently.

what replaces the naive loop isn't a smarter loop. it's usually a simpler one with a harder exit condition.

[miros @m1rossgm]

task in reads codebase writes code runs tests reads stderr patches re-runs. no human touching any step in between.

everyone's reviewing claude code as if it's a better copilot.

it's not that.

18 months ago that was a demo you'd screenshot.

the part nobody's really clocking is that the 'coding assistant' framing is doing real damage to how people evaluate it. they compare response quality when they should be watching what happens when you hand it a failing test suite and walk away.

watch what actually happens on screen. it's a different category of tool.

[promptlab @promptlab11]

ok the “trying to get a rare AI pull” trend has me feral.

people are treating AI generation like a loot box now.

roll, reroll, roll again, chasing the one output that breaks the pattern.

we spent years gambling on gacha games for that same dopamine hit.

now we’re doing it with prompts instead of skins or cards.

same brain chemistry, brand new interface.

nobody’s forcing this, we’re opting in for fun, which is the scary part.

show the grind, tease the wait, drop the rare one at the end.

AI didn’t invent this loop, it just gave us infinite pulls.

save this, this format is about to be everywhere.

[aibalakay @aibalakay]

people keep asking “when will AI actually be useful” like it’s still a hypothetical.🤖

it’s not a hypothetical anymore. it’s a Tuesday. it’s a random 30 second clip nobody outside a small circle even noticed.

the gap isn’t between “AI works” and “AI doesn’t work.” the gap is between people who checked this week and people still running on what they saw six months ago.

one of those groups is about to get blindsided. it’s not the first one.🔥👇

[miros @m1rossgm]

the linkedin profile scrape into a sheet every morning.
someone opens a browser, finds 10 linkedin profiles matching a search, pulls job title, company, headcount, and pastes it into a sheet. thirty minutes of work. paid $15/hr. done by a human.

now: one mcp connector to a search source, one to google sheets, claude sitting in the middle reading the prompt, calling the tools, deciding when it's done.

one prompt in → 10 rows populated → no human touching any step in between.

that's not the 'claude runs a company' demo. that's a repetitive cognitive task that used to require a hire, now running on a cron job.

the part nobody's really clocking: mcp didn't make this possible by being smarter. it made it possible by giving the model hands. read a page, write a cell, loop until done. that's the primitive.

you don't replace a task by building a chatbot. you replace it by pointing a pipeline at the tools the task used to touch. the connector layer is what changed. the model just needed somewhere to reach.