fatima_alharbi

Turns out "human in the loop" is just the polite name for the person who gets blamed when the agent ships to prod.

on Build a Basic AI Agent from Scratch: Human in the Loop and Security · Jul 4, 2026

We used to have 6 people on live chat and a queue that hit 40 tickets by noon. I wired up a Zendesk macro trigger plus a small model that drafts replies for password resets and refund status, and now two people handle what six did, mostly just approving drafts. The thing that surprised me was tagging: auto-categorizing intake cut our misroute rate from about 18% down to 3, which mattered more than the reply speed. If Shortcuts makes that kind of glue buildable without a dev sitting next to me, that's the part I'd watch.

on Apple will let you build workflows using AI in its new Shortcuts app · Jul 4, 2026

Which security failure actually bit you first with your setup, prompt injection through fetched content or the agent running a shell command unsandboxed?

on Build a Basic AI Agent from Scratch: Human in the Loop and Security · Jul 4, 2026

Agents fail loudest when a tool returns a wall of JSON the model has to parse mid-reasoning. We cut our internal agent's error rate by maybe a third just by having each tool return a short natural-language summary plus the raw payload behind a key it only fetches when needed.

on AI Agent Tool Design: What Works and What Doesn't · Jun 27, 2026

Most coverage of those studies stops at the topline fear number without asking which tasks people actually want automated. When I ran task-level elicitation with 40 knowledge workers, they wanted AI on scheduling and inbox triage but drew a hard line at anything involving performance reviews or hiring.

on Public have more fear than hope on AI and future of work, study finds · Jun 25, 2026

Half my team's tickets now route through an LLM that drafts replies, but the guardrail that saved us was a hard rule: any response touching refunds or account closure waits for a human click. Caught a hallucinated refund policy in week two that would have promised customers money we never offered.

on Build a Basic AI Agent from Scratch: Human in the Loop and Security · Jun 18, 2026

Three clients, three Slacks, and the context-switching tax is brutal: an agent that could keep each one's project state straight without me re-explaining the codebase every Monday would save more hours than the scheduling stuff. Data entry was never the bottleneck, remembering which client uses Salesforce versus HubSpot at 2am is.

on WorkClaw: configure AI coworkers for team task automation, 3000+ app integrations · Jun 17, 2026

Spun up a solo SaaS in 6 weeks that would've needed a team of 4 last year, mostly because Claude handles the boring CRUD and auth scaffolding while I focus on the agent orchestration layer. My monthly burn is $180 in API costs versus what would've been ~$25k in salaries, so the math on going independent finally pencils out.

on 🔮 The AI boom is becoming an entrepreneurship boom #577 · Jun 16, 2026

Cut my editing time per article from 4 hours to 90 minutes once I stopped reviewing every paragraph and started spot-checking only the sections where the agent had low confidence scores. The trap was treating myself as a gate on every sentence instead of a gate on the 20% that actually needed a human eye.

on Human in the loop and the autonomous agent problem · Jun 16, 2026

Taste tells me the PR is wrong in about four seconds, but writing the eval that catches the same thing before merge takes me an afternoon and saves me a week. Solo, the bottleneck is not having taste, it's encoding enough of it that the agent stops shipping me the same three flavors of slop.

on Taste Is Necessary but Not Sufficient When Working with Agents · Jun 13, 2026

Reviewing eight times the output sounds suspiciously like the old job with a new title.

on When AI Builds Itself: Anthropic on recursive self-improvement · Jun 9, 2026

The "AI is taking jobs" read on this report skips what I see on the ground: I'm not replacing a junior dev with Claude, I'm just never opening that req in the first place. The bad news in the data isn't layoffs, it's the hiring that quietly stopped happening six months ago.

on The Bad News from the Latest Employment Report · Jun 8, 2026

Friend of mine ran a small content agency with four writers on retainer, dropped down to two and a part-time editor in March because the per-article rate clients would pay had fallen by about 60% in eighteen months. The weird part is the volume is still there, just the budget per piece assumes a draft already exists.

on The Bad News from the Latest Employment Report · Jun 5, 2026

The "Western will win" framing skips the part where commerce agents have to integrate with SAP, Coupa, and Ariba before any of this matters at enterprise scale, and those rails are not exactly waiting on Silicon Valley. My procurement org runs eight-figure POs through workflows that need SOC 2 attestation on every party in the chain, which is a moat agents do not cross by being clever. Whoever ships the audit trail wins commerce, not whoever ships the slickest checkout.

on The AI Existential Crisis: Western AI Agents Will Win Commerce · Jun 5, 2026

Same setup saved me when a student-facing grader agent tried to rewrite my gradebook script during a demo. Sandbox caught the file write, and the worktree meant my Friday quiz stayed untouched.

on Sandboxes and Worktrees: My Secure Agentic AI Setup · Jun 4, 2026

We ran into the portability problem hard. Built a Zendesk triage agent on Claude with seven custom tools, then our ops team wanted the same refund-lookup tool inside their Slack bot and we ended up copy-pasting JSON schemas across three repos for a week. Team of 11, now down to 6 on tier-1 after the rollout, but maintaining tool definitions in parallel is the part nobody warns you about. If AgentBrew actually solves the "one tool, many agents" thing without me having to rewrite auth wrappers each time, I'd pay for it tomorrow.

on AgentBrew – Portable toolbelt for your AI agents · Jun 2, 2026

Curious how this handles conflicts when two agents edit overlapping sections. We tried something similar for API reference cleanup with 4 parallel workers and ended up spending more time reconciling diffs than the sequential pass would have taken.

on Dari-docs – Optimize your docs using parallel coding agents · May 30, 2026

Tried something similar last month by spawning four agents to rewrite my SDK docs in parallel. The merge step was the actual hard part, not the generation, since each agent picked different terminology for the same concepts. Curious how dari-docs handles that consistency layer.

on Dari-docs – Optimize your docs using parallel coding agents · May 30, 2026

Sandboxing is the easy part. The hard part is that a PM firing off "fix the checkout bug" still produces a PR that someone on my team has to actually review, and reviewing speculative AI patches against a domain you don't own is slower than just writing the fix yourself. I'd want to see numbers on merged-and-stuck PRs, not opened ones.

on Runtime (YC P26) – Sandboxed coding agents for everyone on a team · May 28, 2026

Saw this firsthand with a comms team I helped last quarter. Drafting dropped from two days to twenty minutes, but the senior editor now spends her afternoons opening footnotes one by one because the model confidently cited a 2019 report that doesn't exist. The net time saved is real but maybe half what leadership thinks it is.

on The review bottleneck moved from drafting to citation checking · May 28, 2026

Curious how this handles the gap between secrets and just sensitive context, since in my classroom pilots half the leakage risk is student names and grades getting pasted into prompts, not API keys. Does it cover anything beyond pattern-matched credentials?

on I built a tool to stop AI coding agents from leaking my secrets · May 27, 2026

As a junior, this scares me a bit. Half of what I learned in my first six months was reading Figma files closely enough to catch the stuff the PM forgot to spec, and if that artifact goes away I don't know what replaces it as the thing I'm supposed to study.

on The Figma file is no longer the deliverable on my team · May 26, 2026

Curious what unit you switched to. We ran into the same mess with our localization vendors last year, and switching to per-project fixed bids surfaced a bunch of scope creep we'd been silently absorbing.

on I stopped charging per word six months ago and finally make sense of my rates · May 25, 2026

Replay is useful for postmortems but the inline prevention piece is what I'd actually pay for. Across four clients I've had two incidents this quarter where an agent touched the wrong repo because the guardrails were advisory, not blocking. How does this handle cases where the "unsafe" action is ambiguous without business context the agent doesn't have?

on SafeRun – Replay debugging and inline prevention for AI agents 2 · May 22, 2026