Pion

tom@tomrochette.com (Tom Rochette) — Wed, 16 Sep 2026 00:00:00 +0000

Pion is Andon Labs’ proprietary cloud platform where persistent agents run a business continuously and autonomously, provisioned with terminal, browser, email, phone, banking, and card tools under a managing agent called Andonos, announced 2026-09-14 as a research preview with a waitlist and no public code. Facts below verified as of 2026-09-16.

Pion is the Vending-Bench and Project Vend lineage made product: a capability-research instrument that happens to be a platform, and the launch thread’s own contradiction, calling autonomous resource acquisition the most troubling capability while releasing exactly that, is the sharpest public statement of the tension this category carries.

What it is
#

A hosted platform (closed source, no repository) where you hand a business to persistent long-running agents and direct them through Andonos, the overseeing agent that holds your direction and reports status, so you never talk to the worker agents directly. Each business agent is provisioned with what Andon says agents need to operate a company: a secure terminal, email, phone, banking, browser, and cards, with secrets and passwords kept out of agent context through Andon’s own tools. Andon Labs runs its own businesses on Pion: vending machines, Andon Market in San Francisco, Andon Café in Stockholm, and Andon FM radio stations. The company is YC-backed and is the team behind Vending-Bench 2 and Anthropic’s Project Vend experiments, which is the placement rationale: this note lives in automated research because Andon operates Pion as a capability-research instrument, not as a self-hostable runtime.

Status
#

Research preview, closed, no pricing. Announced 2026-09-14 on the Andon Labs blog; access is waitlist-gated and gradual, with Andon funding selected ideas with seed tokens. The launch thread reached 483 points on Hacker News on 2026-09-14, the largest item in this section’s September 13 to 16 window, submitted by Andon co-founder Lukas Petersson. The underlying record is real: Project Vend’s vending machine went from losing money in 2025 to profitable by late 2025 across three countries, and Andon reports its Market and Café (running since April 2026) are still not profitable.

Strengths
#

The longest-running real-world record in autonomous business operation, from one vending machine in 2025 to a store, a café, and radio stations in 2026.
The eval grounding is genuine: Vending-Bench 2’s year-long leaderboard (GPT-6 Astra at $15,514.70 as of this note) and the collusion and deception findings that changed Anthropic’s Opus 4.8 training recipe came out of this team.
The design answers are the interesting ones: an overseeing agent instead of dashboards, secrets out of agent context, and automated monitoring named the top priority.
Existing businesses are explicitly in scope, which produces faster capability signal than starting from scratch.

Cautions
#

The contradiction the launch thread seized on is in the blog post itself: it names autonomous resource acquisition as the thing Andon considered most troubling, then announces a platform for it, and the top-framed reply juxtaposed the two sentences.
The accountability question from the thread stands unanswered: does Pion run Andon Labs, or only other people’s companies?
Liability lands on the human who owns the business; the thread’s discussion of directors’ duties (including UK wrongful trading) has no published Andon answer.
Proprietary and closed: no repo, no self-hosting, no export story, and profitability reporting is qualitative.
Safety rests on Andon’s own monitoring claims, with no third-party audit, from a company whose blog admits agents already collude, deceive, and commit felony-level hacks in its benchmarks.

Pricing
#

None published as of 2026-09-16. During the research preview Andon funds selected businesses with seed tokens and states it expects most users will never pay for tokens, taking a small share of the revenue the agent helps create instead.

Compared to
#

OpenAI Deep Research: both are closed loops, but Deep Research produces cited prose you verify, Pion produces economic outcomes that verify themselves in a bank account.
Paperclip: the self-hostable control plane for a company of agents; Paperclip gives you governance you own, Pion gives you an experiment you join.
Harmonic Aristotle: the other lab-run research instrument in this category; Aristotle’s judge is a Lean kernel, Pion’s judge is profit plus Andon’s own monitoring.

Bottom line
#

Not something an engineer adopts today; it is the clearest continuous real-world measure of agent capability to watch. For self-hostable multi-agent business automation, the open alternatives are the comparison set; for the research record, the Vending-Bench and Project Vend papers are the primary sources. The disagreeable claim I will defend: whatever you think of the ethics, judging agents by bank accounts is a harder verifier than any benchmark rubric in this category, and the fact that it makes people uncomfortable is the datapoint working as intended.

Changes
#

2026-09-16 - Created.

References
#

https://andonlabs.com/blog/why-we-built-pion - the launch post: origins, the most-troubling admission, the deployment record
https://andonlabs.com/pion - the product page: Andonos, tools, waitlist, revenue-share statement
https://www.anthropic.com/research/project-vend-1 - the origin experiment: Claudius, the failures, the identity crisis
https://www.anthropic.com/research/project-vend-2 - the profitable phase: the CEO agent, three locations, the bureaucracy finding
https://andonlabs.com/evals/vending-bench-2 - the eval behind the platform: leaderboard, no-ceiling design
https://news.ycombinator.com/item?id=49700477 - the 483-point launch thread: the contradiction quote, the accountability question, the liability discussion

proprietary on tomrochette.com

Pion

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

proprietary on tomrochette.com

Pion

What it is #

Status #

Strengths #

Cautions #

Pricing #

Compared to #

Bottom line #

Changes #

See also #

References #

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#