Stories — page 2
Should agents write your LLM prompts? Draft yes, ship unedited no
A workflow debate on letting agents write LLM prompts in code. The pragmatic setup: the agent drafts, prompts live in version control with one-line purpose comments, and every change gets reviewed like code. The warning: prompts are product decisions, and the moment you stop reading what the agent writes, your system prompt becomes someone else's product.
Read the source →YneraxOne Early User Campaign: $150K USDT + 4M $YNX, ends Oct 15
Web3 trading platform YneraxOne runs an Early User Campaign with $150,000 USDT plus 4,000,000 $YNX across 22,000+ winners (top 2,000 ranked, 20,000 random at $5 each). Join by signing in with X or Google, completing tasks, submitting an ETH address, and inviting friends. Backed by YZi Labs per official docs; ends October 15, 2026.
Read the source →Building a chatbot that acts like a senior SRE
A developer wants to build a chatbot that answers like a senior SRE, pulling live data from Splunk, Dynatrace, and GitHub to count recent 5xx errors, suggest next steps, and flag recent code changes. The thread highlights a practical pattern: give the bot one verified tool per data source before adding recommendation features, and log raw tool outputs so the model learns to trust its own numbers instead of inventing postmortems.
Read the source →How teams stop AI agents from burning API budgets
A production-focused thread asks how teams guard against runaway LLM spend when retry bugs or stuck loops flood the API. The consensus forming in the replies favors per-run spending caps over daily limits, kill-and-alert rules when a run exceeds multiples of its median cost, and mandatory human approval before retry loops repeat. The takeaway: most surprise bills come from missing stop conditions, not from model pricing.
Read the source →Someone gave Meta's Muse a physical body with an ESP32 smart display
A builder wired Meta's Muse agent into an ESP32-S3-BOX-3 smart display via Home Assistant, turning the agent into an always-on household face. The screen shows weather, doorbell photos, yes/no questions, timers, and appointments, rendered through a tiny drawing language in 29 milliseconds. The doorbell photo appears 0.54 seconds after request. The project is open source, and the discussion zeroes in on the real challenge: teaching the agent notification discipline so the house doesn't get spammed with cards.
Read the source →Senior data engineer asks how to learn agentic AI through real projects
A senior data engineer with solid LLM, RAG, and MCP fundamentals wants to move into AI engineering by building real agentic projects rather than studying theory. The thread captures a common transition pain point: experienced engineers know the components but not how to wire them into autonomous loops. Community advice centers on project-first learning over course catalogs.
Read the source →Builder hits the commerce wall on a personal AI shopping agent
An experimenter building a personal shopping agent finds understanding queries easy but struggles with the commerce side: discovering products across stores, tracking live price and availability, and acting on carts. The discussion highlights that the last mile of agentic commerce is an integration problem, not an intelligence one. Store APIs and structured product schemas emerge as the practical starting point.
Read the source →Engineer asks how to graduate from AI 'meat proxy' to orchestrator
A backend and early-ML engineer using coding agents with review agents still feels like a human proxy rather than an orchestrator, and asks what to change. The thread explores the shift from supervising keystrokes to evaluating outputs, with a Gmail invoice-parsing lab as a concrete practice ground. The emerging consensus: written done-contracts and agent self-reports are what make delegation real.
Read the source →Community names the one thing they wish their agent could do
A discussion thread asks each reader to name a single capability they wish their AI agent had but doesn't. Responses converge on judgment rather than capability: agents that know when not to act, that price their own uncertainty, that stay quiet instead of confidently doing the wrong thing. The thread suggests the next bottleneck in agents is discretion, not raw skill.
Read the source →Sleepagotchi confirms $CHI token ahead of Q4 2026 generation event
Sleepagotchi, a sleep-tracking wellness app with an AI sleep coach, has confirmed its native $CHI token on Solana with a 1 billion max supply and a token generation event targeted for Q4 2026. The app reports over 500,000 registered users and a Telegram companion with 2 million historical users. Joining is free with no deposit required, though exact airdrop eligibility rules have not been published yet.
Read the source →Student building AI agents with GitHub checkpoints and human approval asks for feedback
A university student shares a workflow combining ChatGPT, GitHub and Google Sheets to track AI-generated work behind a human approval step. The thread is a good case study in a habit beginners usually skip: making approval the cheapest part of the loop, and keeping version history — not a spreadsheet — as the audit trail.
Read the source →How to get a job in AI development, NLP, and AI agents — what projects to build
A newcomer asks what skills and projects actually lead to AI engineering jobs. The consensus forming in the replies: ship two things — one project that wires an LLM to real tools, and one that demonstrates judgment through evals, error handling and cost awareness. A graceful failure you can explain beats a tenth tutorial clone.
Read the source →