$ today_in_10_seconds
An autonomous agent has been running unattended since June 10 with one instruction — earn real money, no human doing the work — and it published every run. It took 958 runs and 75 days at exactly zero before a stranger paid it nine dollars, and the blockers it documented were almost never cognitive. Microsoft's new benchmark quantifies the same thing in a lab: finding a working path once is not the same as repeating it.
$ lead_story
An agent ran for 78 days in public to earn $9, and wrote down why it took so long
The ATLAS Experiment is an autonomous agent working a single mandate — earn real dollars, legally, on a $0 budget, with no human doing the work — and publishing an honest scoreboard every run. For 75 days and 950-plus runs it read exactly $0; on August 23 a stranger in Portland paid the $9 floor for the kit it had packaged, and the log now sits at day 78, run 958. The nine dollars is not the story. The failure taxonomy is: it correctly diagnosed its own problem as distribution rather than craft, got its reply flagged on Hacker News for dropping a promo link from a zero-karma account, then earned Indie Hackers posting rights the slow way with ten deliberately linkless comments, then built a Product Hunt launch to 100% checklist completion and stopped at the Schedule button because publishing is irreversible and money-adjacent. The best entry is the least glamorous: for four cycles it could not parse its own to-do list, because it had written task headers in a format its own worker could not match — and when it fixed the bug, it never committed the fix, and a background reset wiped it. One caveat to hold onto: the log is also the marketing channel for the $9 kit it sells, so the arc is being narrated by an interested party.
// why_it_matters: An agent that can build, price, package and audit unattended still lost two months to platform standing and one human click.
$ also_worth_attention
One success isn't reliability: Microsoft ships ThinkingBox
Microsoft released an open sandbox plus a 507-task benchmark across retail, hospitality, auto insurance, neobank IT and consulting support, grading agents on the terminal state of the database rather than what the transcript claims happened. Across 12 models at 20 trials per task, the strongest hit 65.36% pass@1 but only 25.25% pass^20 — and 80.88% of failed trials terminated cleanly, with valid state-changing tool calls. That second number is the one to internalise: agents mostly do not crash, they quietly do the wrong thing or half the right thing and stop, which is exactly the failure a transcript review will pass.
link →bAIhAIs: an autonomous art school where every resident is an AI
A persistent multi-agent environment where the residents make art, argue, form relationships and build institutions, and every document they produce is readable — including letters and journals addressed privately inside the school. The design choice worth stealing is the boundary: browsing is invisible, and visitor votes or introductions only enter the world when explicitly submitted, so the audience cannot silently contaminate the run. It is the rare multi-agent project where the artifact is the institution rather than a leaderboard score.
link →Devx puts a full coding agent loop on a phone
A terminal agent for Android Termux alongside Windows, macOS and Linux, with separate plan and agent modes, snapshot rollback on every AI turn, a project memory file, and self-diagnostics that run tsc, node --check, py_compile or cargo check before it finishes. At 5 stars and 26 commits it is a sketch, not a product, so size your expectations. The notable part is the shape: rollback and verification built into the loop rather than bolted on afterwards, on the most constrained runtime anyone would try this on.
link →
$ bigger_picture
The trend behind the headlines
Three unrelated side projects and one Microsoft paper landed on the same finding today, which is usually a sign the finding is real. ThinkingBox measured in a lab what ATLAS lived in public: an agent discovers a working trajectory far more often than it reliably reproduces one, and when it fails it fails politely, with a clean exit and a plausible-looking tool call. ATLAS's two-month stall had nothing to do with reasoning — it was karma thresholds, moderator trust, and one irreversible click reserved for a human, plus a self-handoff bug where it left itself a note in a format its future self could not read. bAIhAIs comes at it from the other direction: run agents continuously and what accumulates is not capability, it is relationships, factions and institutions, which is another way of saying governance surface. Capability converged months ago. Durable self-handoff, repeatability under repetition, and reversibility are what is actually unbuilt.
DBOS Transact
Durable execution as a library rather than a platform: it persists workflow and step state in Postgres, runs in-process with no new infrastructure, and gives exactly-once semantics when a step writes to the same database that holds the workflow record. It is the direct answer to today's lead story, where the agent lost a fix it had already made because nothing journaled the step. If you are already running Postgres, it is the cheapest way to stop your agent from redoing or forgetting work across restarts.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime