AgentSky Review

Two meters, both published. Model tokens at provider list prices with zero markup, compute at $0.117 to $1.066 per awake hour, $0.331 on the default machine. No seats, no subscription, no minimum. That is a pilot finance can approve in a meeting.

What you're buying is optionality. Eight harnesses, eleven models, and switching either is an argument in a request rather than a rebuild. Three named customers describe agents their own users touch, at 150,000 and 250,000 agents launched. Vendor-stated, but specific enough to check.

The exposure is upstream. Every harness here is somebody else's open-source project and every model somebody else's API. If a provider reprices, so does your invoice. Run one production workload for a quarter and watch what the compute meter actually does before moving anything critical.

The interesting decision is where the abstraction sits. A model gateway swaps weights behind one endpoint; this swaps the whole harness — Claude Code, Codex, Hermes, opencode — while keeping the session, the tools and the channel intact. If we build on it, our dependency becomes the request shape, not any single agent project.

Architecturally the state story carries the weight. Each agent gets a persistent machine that suspends when idle and wakes on the next message, so crash handling and reconnects sit on their side of the line. Add-ons ride the same call: browser, search, transcription, roughly 2,000 app connectors.

The three-year question is upstream drift. Harnesses evolve independently and a benchmark spanning May to August 2026 is a snapshot, not a trend.

Same machine, two very different invoices. The default 2 vCPU / 4 GB bills $0.331 an hour. At the vendor's estimate — 20 turns a day, two hours awake — that is $19.87 a month against $78.00 of model spend. Never suspend it and compute alone is $238.46.

That is $97.87 a month for one agent, $11,744 a year for ten. Model tokens run at each provider's list rate, $0.15 per million input on GLM-5.3-Flash up to $10.00 on GPT-6 Astra. Compute spans $0.117 to $1.066 an hour.

The estimator counts two meters. Capabilities are a third — browser at $0.03 a minute, connector calls at $0.029, image generation at $0.422 an asset. Bring Your Own Subscription zeroes the model meter on eligible agents. But top-ups still carry Stripe's 2.9% plus $0.30, and nothing here is contractual: the bill tracks your suspend discipline.

One command clones the agent you already run locally — instructions, model, MCP servers — into the cloud. That's the tell. Whoever built it moves agents between machines often enough to hate doing it by hand.

Day-to-day the shape is good. Sessions persist between messages, suspend on idle and wake on the next one, so you're not writing reconnect logic. Usage is itemised to the turn across both meters, which is what you need at 2am when one agent's bill looks wrong. Channels are parameters, so putting the same agent in Slack and behind an HTTP endpoint isn't two integrations.

What's missing is failure detail. Restart behaviour is documented as managed recovery to the latest durable point, not zero-loss, but retry and timeout semantics are absent, and there is no changelog. You'll find those boundaries yourself.

No card is requested at signup, and the playground is enough to run something real rather than a hello-world. The playground lets you pick a harness and a model from a dropdown and start, and you can hand it a GitHub repo or a forwarded email instead of a toy prompt. That's a better first ten minutes than most developer tools manage.

The billing page does something I wish more products did: it tells you the card fee. $100 of credit costs $103.30. Nobody hides that well and this one just says it.

What nags is the two-meter model. Tokens and awake seconds are both perfectly reasonable and together they're two things to keep in your head. The estimator helps. It's still arithmetic you didn't have before.

Credit where it's due. Zero markup on models, per-second compute, the Stripe fee disclosed. Vendors who plan to make money on opacity don't publish that.

What's absent is louder. No security page. No SLA. The data-handling answer is two sentences long — state kept to run and recover an agent, zero retention only on enterprise plans — and there is no security page or SLA behind it. The customer numbers — 150,000 agents launched, 250,000 — are vendor-stated with no way to check them, and 10K+ sessions in production is a phrase, not a metric.

Exit is fine, which surprised me. The harnesses are other people's open-source projects and the models are public APIs, so leaving means running what you already run somewhere else. That's real portability. The lock-in is convenience, not format.