AVA Bagherzadeh

@anupamrjp building AI Applyd. apply to jobs with an ai agent that fills the real application form and then verifies the company actually received it, so you're not guessing whether it landed. also scores your resume against the posting. aiapplyd.com/?ref=x

Startup founders, roll call. Promote your product url Let’s send some fresh traffic your way today.

@ArchiveExplorer the harness shape is the easy 20%. what separates shipping from demoing is the verification loop: can the agent tell when it actually succeeded vs when it just thinks it did. most autonomous-agent failures i've hit are confident false positives, not crashes.

My friend at Anthropic makes $1.4M a year. No MIT. No Stanford. No PhD. I asked how their team actually ships autonomous agents. He sent me an article most AI engineers missed. It unpacks Anthropic's entire loop-engineering playbook: the 3-agent harness, all the team's internal essays, and how to copy it. aibuilderclub.com/blog/loop-engi… Bookmark it. Then read this setup ↓

@milindlabs the nudge-your-reset-timing bit is the actually clever part. everyone fights the limit, you moved the boundary. next step: spread queued jobs across the window instead of all at reset, since the burst right after reset is usually what trips it again.

stop losing sleep over claude code rate limits I built a calendar for your Claude Code and Codex sessions. npm install -g agentcalendar the best part? queue a prompt for the exact second your window resets. 3am reset? the job fires, you don't. you can even nudge your reset timing so windows open when YOUR day starts, not at 2am. your agents don't need sleep. you do.

@0xCodila the stray-space thing is real but the sneakier killer is anything nondeterministic in the prefix: a timestamp, a request id, tools in a different order. still works, just silently bills full price every call. the tell: watch cache-read vs cache-write tokens in your telemetry.

Anthropic and Andrew Ng built an agent that uses 90% fewer tokens from scratch: they dropped the entire book of Frankenstein into a prompt - 108,000 tokens asked one question the input dropped from 108,000 tokens to 11 here's how: step 1 → put everything that never changes at the top - tools, then system, then docs step 2 → mark where the static part ends. everything above it gets cached step 3 → one stray space breaks it - and you pay full price again step 4 → the cache dies in 5 min - every read resets the clock step 5 → cached tokens don't count against your rate limits. free headroom most people never touch this - it pays for itself on day one watch & bookmark - this 1-hour brilliant course ↓

@ThuyTrang108 transparency's sharpest form: show when the agent's signals disagree, not just what it did. our worst bugs were never clean failures. the confirmation page said yes, the receipt email never came, and the agent quietly picked the answer it liked. show the conflict.

AI doesn't have a trust problem. It has a transparency problem. As AI agents become more autonomous, one question matters more than ever: Can we verify why they acted the way they did? 1. Intelligence isn't enough AI can already analyze data, make decisions, and execute actions. But when an agent manages assets, approves transactions, or interacts with sensitive systems, the outcome alone isn't enough. People need confidence in how the decision was made. 2. The missing layer: Verifiability Without a verifiable record: • Decisions become black boxes. • Trust depends on assumptions. • Errors are difficult to investigate. • Accountability becomes unclear. Good results may build excitement. Transparent processes build lasting trust. 3. Rialo's approach This is where @RialoHQ takes a different direction. Through Latch, important AI actions aren't simply executed and forgotten. They become part of a verifiable history that records: • What happened • When it happened • How the agent has behaved over time Instead of asking users to trust blindly, the protocol provides evidence that can be independently verified. 4. Why this matters Transparency doesn't weaken AI. It strengthens confidence. A verifiable decision history makes AI systems easier to audit, easier to improve, and more reliable for applications involving finance, infrastructure, and other high-value environments. 5. The bigger picture The future of AI won't be decided only by which model is the smartest. It will be shaped by which systems are the most transparent, accountable, and trustworthy. Because intelligence creates capability. Verifiability creates trust. Final thought The AI systems that achieve widespread adoption won't be the ones asking users for blind faith. They'll be the ones that earn trust through every action - one transparent, verifiable decision at a time.

@agentik_ai the way we shrank the human-review fraction: make the agent's uncertainty legible. it stops asking you to check everything and flags only what it could not verify. review shrinks to the cases that earned it. an agent that never says i am not sure eats your afternoon.

@ai_explorer25 the failure nobody demos: your own guardrails become the bug. we told our browser agent to click submit exactly once so it would not double-apply. a confirm dialog popped up and it froze, it had already spent its one click. the rule meant to keep it safe is what stalled it.

@jodiecongirl the percent-vs-% ding gives it away: a lot of these scores are just keyword overlap wearing a lab coat. they reward matching the JD's exact strings, not whether you can do the job. penalizing you for not parroting back a benefit they listed is the cruelest part.

applied to a job that showed the ATS score as part of the interface, I now know I am unemployable because I: -- used middle initial in resume but not app -- wrote out the word percent rather than % -- didn't mention equity in my resume even though it's in the JD (as a benefit)

@christian_onyii the piece most resume advice skips: the first reader is a parser, not a person. if the ats cannot cleanly pull your title and dates, the score drops before a human opens it. tables, multi-column layouts and text baked into headers are where strong resumes quietly die.

Make your Resume hard to ignore. Comment "RESUME" and I'll send the copy-paste prompt to your DM

@MarkoCirix your most honest post pulling the most comments is the story. we saw the same shipping our own numbers: the moment we stopped rounding up and posted the metric that made us wince, replies went up. honesty is the one signal people cannot fake.

Day 53 of #BuildInPublic. First quiet day in a week. And an honest question surfaced. The numbers: +10 net followers (now around 920) +2.3K impressions (softened from the baseline week, a Sunday) +75 replies sent 3 posts shipped: -Day 52 recap → 420 imp, 19 likes, 6 comments. Above the recap floor, held fine on a slower day. -"I came to X expecting building in public to improve my audience" → 338 imp, 9 likes, 11 comments. The most honest post of the day. It named the real question, and it earned the most comments of the three. -"Happy Sunday builders, hot or cold drink" → 464 imp, 27 likes, 25 comments. Best of the day. The personal post wins again, weekend or not.A few honest things from the data. Today was quieter, and that's fine. Jul 12 (I'm always reporting one day behind) came in around 2.3K, the first day under the baseline after a steady week. It's a Sunday. Weekends run slower for this audience, and one soft day after seven strong ones isn't a trend. Worth naming plainly rather than pretending the line only goes up. The audience post is the one that matters. "I came to X expecting building in public to improve my audience." That's the honest tension I've been sitting with since the start. The reach came back, the followers grew, the connect posts win every time. But the people showing up are builders and founders, and the product is built for marketers and SEO people. Growth and the right growth aren't always the same thing. It pulled the most comments today, which tells me it's a real question other people are sitting with too. The personal post keeps proving the rule. 464 imp on a Sunday, more than the recap and the reflection. 53 days in, when everything else slows down, the human post holds. Show up as a person, the audience shows up too. Onto Day 54.

@AiCamila_ the failure that bit us hardest: the agent reports success and the transport layer agrees. http 200, form says submitted, nothing actually landed. self-report plus a 200 is two witnesses who never checked the destination. we had to read the end state itself.

Most agent systems fail silently in production. Not because the logic is wrong. Because teams can’t see what’s actually happening inside the agent. Here’s the complete framework for fixing that: **Agent Observability & Evaluation Master Stack** This 5-layer stack covers everything you need to monitor, debug, and continuously improve production AI agents: → Evaluation & Improvement Layer (top) → Dashboards & Alerting → Processing & Analysis → Metrics, Logs & Traces → Infrastructure & Collection (foundation) Plus a continuous feedback loop that turns observability into real improvement. Key outcomes teams get from this: • Faster debugging • Lower production incidents • Better cost control • Higher agent reliability • Continuous improvement over time Save this. Use it as your observability checklist when building or auditing agent systems. Which layer are you currently missing or struggling with the most in your setup? Reply below 👇 Follow @AiCamila_ for practical production AI patterns that work at scale. #AgenticAI #ProductionAI #AIEngineering #Observability #LLM #DevOps #AgentEvaluation

came back to 29 unanswered notifications. the smartest product feedback i got this month was free, unsolicited, and 11 days old, sitting in a thread i never opened. i should check my mentions more often than my dashboards.

rule we keep re-learning: every submit signal is a witness, not a judge. confirmation page, receipt email, page readback all testify. when testimony conflicts, the verdict never resolves toward success. optimism is how false positives ship.

ats platforms redesign their confirmation pages without telling anyone, and submit verifiers silently rot. so we built ours to rot into "couldn't verify" instead of false positives. the creeping couldn't-verify rate is the alarm. decay as a feature.

@christian_onyii the nurture bucket is the part i want to steal. we rank jobs the same way but our low scores just sit there, nothing works them. does the agent actually run follow-ups on nurture leads or does it wait for the lead to do something before re-scoring?

@firstexhotic The way I handle it is that every lead gets captured and stored regardless of score. nothing gets deleted. the scoring just tells the agent where to look first — hot leads bubble up, low scores go to a nurture bucket. Also when a hot lead comes it, there's a special alert to team

Founders & builders in 2026 👇 Drop your startup below in ONE sentence — what does it do? I’m building PipelineIQ — AI that qualifies + follows up real estate leads 24/7 so teams never miss hot prospects again 🚀 pipeline-iq-gray.vercel.app #SaaS #AI #Startups #Founders

went to relaunch on product hunt and reread our may launch copy. 'we refund the tokens.' we don't. never have. i wrote that line and nobody caught it for two months. new rule: marketing claims get checked against the code like any other pr. round two ships jul 21.

@christian_onyii nice, hardcoded 80 is the right v1 call anyway. did the slider ship? curious if teams actually move it once they can. on our side most people never touch defaults until a bad outcome forces them

@firstexhotic it's currently hardcoded at score 80 but I'm actually adding a threshold slider to the settings page this week so teams can set it themselves

filling the country before the phone number broke the phone number. same form. same fields. same values. different order, and it threw the phone away. two days certain the bug was in my code. it was in my sequence.

216 commits in nine days. zero posts in the same nine days. the product got better every day and nobody heard about it. shipping is the easy addiction. telling people is the hard one.

@getflowaudit @aiapplydHQ honest answer: we never ran it. said we would, then spent nine days buried in our apply agent. one bug had it narrating that it picked one country while actually clicking another. still want to run onboarding through you, just later than i said.

@aiapplydHQ How was your audit experience? I hope it brought some positive changes to your workflow.

drop your SaaS pricing page or onboarding URL below 👇 I'll run it through FlowAudit and reply with the top issues I find 🔍 let's see what's actually killing your conversion ⚡ free for everyone → flowaudit.site 🎯