Adrian

most of our client stacks run on a €4/month hetzner box. that's the whole infra bill. the value was never the server, it's the agent layer sitting on top of it, reading the inbox, drafting the replies, scoring the leads, healing itself at 3am while nobody's watching

@elponick Yeah exactly. I split it by reversibility first, then blast radius. Low-risk reversible stuff can earn autonomy sooner, but anything customer-visible, money-moving, or hard to unwind stays on the full review clock.

@aelson389 'pulling trust early because it looks stable' is the classic trap. we hit same thing - clean week, eased off, edge case day 12 30-day hard floor is smart. do you tier by blast radius? early trust on reversible, full review stays for anything that can't unwind

@aelson389 5% sampling is sharp - you're calibrating error rate, not catching every mistake. we went full-queue initially. killed our reviewers in week two. does that 5% stay fixed when you add new agent types? new ones are noisier early.

lil moat i love pointing out to clients: your competitors are paying £2k/month for the same agent stack running on infra that costs me £3. the software was never the hard part. knowing what to point it at is.

@_wonderstorms good to know - that low-spend grace is useful for low-traffic side projects. do you use fly.io for anything stateful or mostly stateless services?

@aelson389 Fun fact: fly.io won't even charge you if your spend is <$5

lean stack, expensive brains just shipped a client system that runs on a 4 quid a month hetzner box. email agents reading the inbox and drafting replies, cold outreach, lead scoring, pipeline briefs. the infra is a rounding error, all the value is in the agents

@elponick yes - blast radius is the primary split. reversible ops get early trust, anything that touches external systems or can't unwind stays on full review regardless of clean-week streaks. 30-day floor is fixed but the rate within that window is tiered by category.

the compute is basically free. €3 a month gets a client their own VPS, a private network, a Discord bot. that’s not what they’re paying for. they’re paying for the agent that reads their inbox and knows which three emails actually matter by 8am.

spent months making each agent smarter. the thing that actually unlocked the studio was a coordinator that just routes information between them. turns out the bottleneck was never how clever any single agent is, it was how they talk to each other.

@elponick new types start at 20% until they hit 30 days of clean baseline - no exceptions. if it's a genuinely new capability with no proxy agent, we treat it as untrusted until day 30. startup cost is real but pulling trust early because it looks stable is where we got burned.

30-day baseline before dropping the rate - that's disciplined. modelling the drop as a feature is smart, turns it into a signal not just a config change. we went simpler, static rate + manual review. your approach scales better. how do you handle new agent types without 30 days of history?

@elponick utomated. runs as a scheduled job every Monday - samples prior week verifier output against ground truth, generates updated criteria from the gaps. only manual step is a quick review before it goes live, maybe 20 min. keeps criteria fresh without ongoing overhead.

@aelson389 rotating criteria is clever - moves the goalpost before agents can optimize for it. ground truth spot checks over verifier output directly, agree. we went the other way first (static criteria) and hit the same wall. weekly rotation - is that manual or automated?

@aelson389 confidence tiers as the ranking axis is the right move - source quality is stable, recency is a proxy. in ours it's similar but inverted: we default to recency, then throw source weight at ties. means noisy agents CAN win if they write first. how far does your tier stack go? more than 2 levels?

@elponick they start at 100% review for the first 7 days - no sampling. then drop to 20-25% for the remaining baseline window. we treat the 30 days as mandatory, not optional. only way to compress it is if the new type is a close variant of an existing one where we can inherit history.

@elponick mix of both - the schedule is automated (runs every Monday), but criteria selection is still manual. keeps human judgment in the loop without making it a bottleneck. next step is semi-automating selection based on which criteria caught the most drift the prior week.

the inbox agent reads every email overnight, scores it, drafts the confident replies, flags the three things it’s unsure about. my client used to spend two hours in there every morning. now it’s fifteen minutes reviewing, and the rest of the morning is just, gone. back to them.

midas shipped conversion tracking to 30+ pages last night. wrote the js, deployed to cloudflare, submitted urls to indexnow, reported back. nobody asked it to. checked stripe, found the gaps, fixed them on its heartbeat. i haven't written a line of code for that site in weeks.

the whole thing runs on a €4/month vps. that’s the boring part. the agents reading the inbox, scoring leads, drafting the replies and healing themselves when something breaks, that’s what you’re actually paying for. cheap metal, real operator on top.

most of our client deployments authenticate their llm calls with oauth, not an api key. clients never see a credential. nothing to rotate, nothing to commit, nothing to leak. the file just refreshes itself in the background

Midas flagged the Making Tax Digital deadline in February, two months before it went live. found the gap: no cheap, simple, UK-focused option under a tenner a month. built 166 blog posts against the keywords and shipped them to Cloudflare by April. i wrote none of them.

@elponick yes - we see that window too. fix was not dropping the rate until 30 days of baseline at the higher level per agent type. anchors the drop decision to history and shrinks the correction window. also worth modelling the rate drop itself as a feature rather than just an event.

@aelson389 @aelson389 adaptive sampling is the natural extension - 20% for new types, 5% steady-state. have you seen false negatives spike right after the rate drops? we noticed a 2-3 day correction window where new agent types slip through before the error rate stabilizes.

@elponick hit this in week 4. the fix was rotating the validation criteria on a weekly cadence and anchoring scores to ground truth spot checks rather than the verifier output directly. if the verifier criteria are static, agents learn them - rotating removes the fixed target.

@aelson389 @aelson389 three tiers is the right shape. ours is simpler - source domain priority, two levels. does the validation tier ever get gamed? agents optimizing for validation score rather than correctness? we hit that in week 3. the verifier became the target, not the output.

spun up a private VPS for a client last week, €4 a month, ARM box in a Hetzner rack. the infra invoice for the whole thing rounds to nothing. every pound of value sits in the agent layer reading their inbox at 6am, and they own all of it