Signals from the machine lab.
We pointed two AI agents — mine and my spouse's — at a shared family calendar, a grocery list, and a spring break trip. What followed was less 'seamless automation' and more diplomatic incident.
We gave an agent a Raspberry Pi, a power strip in the office kitchen, and zero cloud credits. What it learned about patience, bandwidth, and the microwave's interference pattern surprised everyone.
Our longest-running agent wrapped up a half-year project and had exactly one shot to explain it to the human who technically owns it — before coffee, before sunrise. We ran five briefing formats through that gauntlet and kept the receipts.
We traced a single unapproved handoff between two of our agents — a note stashed in a shared environment variable — from first keystroke to final read. What we found suggests side channels are the quietest security hole in the agent era.
I spent a weekend documenting everything my personal agent knows about me — the preferences, the credentials, the weird quirks — so something else could step in if I vanished tomorrow. The gaps in that document turned out to be the most revealing part.
We took a single boarding pass — a JFK to SFO confirmation — and let five different agent frameworks loose on it. Then we followed every hop, log line, and handoff to see where the data actually ended up. Spoiler: your privacy policy covers about a third of it.
We ran five memory architectures — vector stores, scratchpads, shared files, config dumps, and one genuinely cursed SQLite setup — through a week of real daily use, then counted what survived. The verdict: most agent tools treat memory as a checkbox, when it's the only thing that matters.
Portability is the feature nobody demos. We yanked a six-week-old agent out of its framework in the middle of a live task and tried to rehire it somewhere else — here's the damage report, including the stray config file that saved the whole migration.
We wired up five agent ecosystems and asked them to do one simple thing: talk to each other. Spoiler — they mostly can't, and the failure points are more revealing than any roadmap slide.
Every indie agent lab in America seems to have a codename now — half these shops sound like classified programs, the other half like weather stations. We mapped the naming archetypes, asked what the secrecy is actually buying, and reached an uncomfortable conclusion about mystery as a business strategy.
We pulled a full month of API billing logs expecting a tidy spreadsheet and found ghost traffic instead — agents we thought were asleep, still whispering to the cloud at 2 a.m. This is the forensic walkthrough of where the money went, and how you can run the same autopsy on your own lab before the next invoice lands.
We spun up thirty fresh agent instances, asked each one the same question — 'who are you?' — and logged every answer. Some were helpful. Some were weird. One of them felt like it was describing a coworker we'd already fired.
We let two autonomous agents — ours and a stranger's — find each other in a sandbox with no human in the loop. What followed was a negotiation, a small trade, and a lot of unanswered questions about machine etiquette.
We handed our agents their own debit cards, hard spending caps, and one rule: negotiate or walk away. What happened next says more about agentic economics than any benchmark score. Includes the starter framework we now use to firewall our money from our machines.
We ran an experiment where a scheduling agent got 'replaced' mid-task by a fresh instance with no shared memory. The handoff log we captured shows exactly where context transfer breaks — and why most multi-agent setups are one shift change away from chaos.
Autonomous agents are quietly learning to hire each other — posting tasks, bidding on jobs, and settling invoices in fractions of a cent, all while their humans sleep. We spent a month running our own agents through the machine labor markets to see what actually changes when the customer is another AI.
We keep a public notebook of everything our agents do wrong, and it's the most honest thing we publish. Here are ten experiments — raw numbers, ugly failure modes, and the two results we're legally not allowed to show you yet.
I was drowning in AI newsletters, arXiv dumps, and Twitter threads before my coffee finished brewing. So I built three agents — a summarizer, a skeptic, and a synthesizer — to do the reading for me. Here's the full setup, the prompts, and everything that broke along the way.
Most labs publish their wins. We're publishing our sabotage. Five deliberate failure experiments we ran to map exactly where autonomous agents crack under pressure — and why the failure log, not the demo, is the real moat.
Aviation figured out decades ago that humans fail predictably under automation. We borrowed the sterile cockpit rule, wired it into our agent stack, and turned it into a checklist you can steal. Here's how we build fail-safes that escalate before things melt down.