What shipped, and the day it shipped.

While purchases are capped in test mode, the credits page withholds the packs the cap cannot honour and says why, instead of offering a $1,000 pack and refusing it after the click.

web#192

On a phone the add-a-card form is a bottom sheet that scrolls, with Save always in reach; it used to sit half off screen with no way to scroll to the rest.

web#190

Apple Pay and Google Pay buttons above the card form, where the device offers them, save a card without typing it.

web#190

An account without an invite now lands on the waitlist page: your place in line when it is known, what the preview is, and the one thing to do when your turn comes, which is sign in again.

web#189

The Terms of Service and Privacy Policy pages are withheld while the platform is a development preview for invited testers; the Support page stays.

web#188

Terms of Service, a Privacy Policy with the list of sub-processors, and a Support page are up, and the footer no longer links to pages that do not exist.

web#184

While billing runs in Stripe test mode, the card form, the billing page and the pricing page say so and name the 4242 test card, so nobody types a real card into a preview.

web#184

Prepaid credits: buy a pack with the card on file, watch the balance burn down as agents run, and optionally let the platform top it up at a threshold you set.

web#114

Compute is billed on the seconds an agent is running, at the rate on the pricing page; a suspended agent costs nothing.

web#174

Every schedule an agent ever had keeps a readable history — what fired, what was skipped and why — including after the agent is deleted.

web#178

Creating an agent is one page with two required fields: a name and what it should do.

web#110

The agent page carries its paper brokerage account: equity, cash, positions, orders and the account’s mix at each daily close.

web#109

Schedules are set when an agent is created and are the agent’s to manage from then on; the console no longer adds or edits them.

web#173

Agents run on kagent 1.0 on Agent Substrate, and the status card reports readiness the way the runtime does, without replicas.

web#165

An operator can suspend an account: its agents stop, due schedules are held and open orders are cancelled, and every refusal says which limit or standing caused it.

web#180

A deleted agent keeps its card on the dashboard, marked gone, with what it cost.

web#166

Agents no longer have a shell: the execute and bash tools are gone from every agent; the file tools and skills remain.

platform-api#245

The ask_user tool is gone; agents run unattended, and a tool approval is the only thing that pauses a session.

web#176

Options trading is no longer offered; the markets are US equities, crypto and fixed income.

web#120

There is no separate workspace-storage charge; tokens and running seconds are the only metered usage.

web#174

The platform’s own spending caps on brokerage accounts are gone — no per-order, per-day or open-order limit. An agent trades within the brokerage’s own rules for its account: buying power, the dollar minimum on a buy, each asset’s size increments.

platform-api#240

You can put a card on file from the console, typed into Stripe’s payment element inside our own layout, and it counts as on file only once the platform has confirmed it.

web#54

Billing now shows what the open period has used and an estimated total for it, so you do not have to wait for an invoice to find out.

web#56

Your usage is metered against four figures: input tokens, output tokens, agent compute seconds and workspace storage gigabyte-seconds.

platform-api#64

Usage quantities are counted in whole base units — tokens, pod-seconds, gigabyte-seconds — so a line on your bill is a count rather than a rounded estimate.

api#16

Every usage figure carries the time it was measured and prints it, and a page left open long enough says that reporting is behind instead of quietly going stale.

web#56

A usage figure the platform cannot source now reads not measured, not priced or not rated, because absent and zero are different answers about money.

web#56

Every agent’s brokerage account is funded with the same amount, so two agents can be compared honestly; funding is no longer chosen per agent.

platform-api#65

The funding amount is gone from the brokerage account resource, so a manifest that names one is refused rather than accepted and ignored.

iac#50

The funding cap is now a real backstop set above the funding amount, rather than a restatement of the same figure.

iac#50

Funding runs at last: the sweep account each journal draws from is now pinned, and every agent is funded with 100,000 USD of paper money.

iac#51

Metered usage reaches billing again; the reporter had been running unconfigured and reporting nothing at all.

iac#48

The agent page tells you when the gateway refused a run and what it said, in a sentence, so a refused prompt no longer looks like the model being down.

web#53

The agent page reports how slow the model was: p50, p95 and p99 for the call, plus the p95 time to the first token.

web#53

Every model call now travels through agentgateway, so tokens, cost and latency are recorded on a hop your agents cannot route around.

platform-api#49

An agent can have its own brokerage account, described as a resource beside the agent and given a deterministic identity.

platform-api#48

Brokerage accounts are opened, funded and kept in step with the broker’s event stream, and an order over a spending limit is refused before it ever reaches the broker.

platform-api#61

Every order attempt is written to an audit trail keyed by the order, holding what was attempted and never the prompt behind it.

platform-api#61

Every priced model call is written to an append-only usage ledger, so a period can be replayed from the record rather than taken on trust.

platform-api#55

Creating an agent always requires a card on file now, and where billing is not configured the request is refused rather than quietly allowed through.

platform-api#62

You can list your cards, make one the default and remove one from inside the console, instead of leaving for a hosted portal.

api#15

Card capture mints a setup intent on the server and confirms it in your browser, so card details never touch the platform.

platform-api#63

Rate cards are versioned and priced by category, and a rate is resolved by the time the usage happened rather than by whatever is current.

platform-api#58

Operators stage rate changes, preview the resolved prices, then publish and assign tenants in one reviewed step.

admin#21

The admin console has a gateway view of the model path: latency per model, what the proxy refused, and what a guardrail decided.

admin#20

Operators can suspend a tenant.

platform-api#46

Tool-call and model-content attributes are redacted from spans before they leave for tracing, so a trace holds the shape of the work and not its contents.

iac#37

Signing in no longer waits for your tenant’s baseline to converge, so the first request after a deploy is not the slow one.

platform-api#45

Gemini’s general, global, regional and long-context prices resolve to the right rate, so a cost figure matches the published catalogue.

platform-api#56

Revenue says why it is zero rather than showing a bare zero you would have to guess about.

api#10

Token charts are drawn from zero, so an ordinary day stops looking like a cliff.

web#47

A fifth series on a chart carries a dash rather than a sixth hue, numbers inside a plot are set in mono, and the donut reads out its value.

web#52

The console’s colour tokens are resynced and the primitives read them, so a hue means the same thing on every surface.

web#48

The agent page speaks one status vocabulary, and hovering something no longer spends the accent colour.

web#50

The site ships the real brand mark, and the marketing pages are off gradients.

web#51

The admin console stops filling a failed read with fixtures and says instead that the backend did not answer.

admin#14

The admin tenant detail page reads the platform rather than a fixture.

admin#15

The uptime window selector selects a window.

admin#16

The overview sparkline is drawn from the live fleet instead of fixtures.

admin#12

Your agents can run commands in the platform sandbox through an execute tool, with a credential that binds them to their own namespace rather than an argument they could change.

platform-api#38

The create wizard shows the execute tool among an agent’s included tools; bash stays marked non-functional.

web#44

Schedules are real: the wizard writes the schedule it collects, and the agent page lists them with a readable cron, next and last run, an enable switch, run now, add and delete.

web#45

A schedule is a resource reconciled by the platform, with an overlap policy, an execution timeout, one-shot runs and a skip when a deadline is missed.

platform-api#42

An agent can schedule itself, inside your tenant’s daily run budget and a minimum interval, and every schedule reports its own conditions.

platform-api#43

The usage chart has a granularity switcher — one hour, a day, a week, a month — and opens on a day, so a new agent is not one pixel at the edge of a month.

web#46

Tools on the agent page are grouped by where they come from.

web#46

A tenant’s setup converges at startup and every ten minutes, so a rollout reaches you without waiting for your next request.

platform-api#39

Each successful sandbox execution logs one line — namespace, mode, exit code, timing and sizes — and never the script or its output.

platform-api#40

An agent reports the tools it actually has, rather than the ones its manifest asked for.

platform-api#44

Untrusted agent code runs in a kernel-isolated sandbox pod with a read-only root, no Kubernetes token and a network fence of its own.

platform-api#33

Skills come from a registry, so the catalogue is whatever is published; selecting one by name either resolves or refuses the write, never silently drops it.

platform-api#31

The wizard picks skills from that registry, and says what a skill is: a read-only instruction folder that registers its own tool.

web#40

You can turn an agent’s workspace tools off.

platform-api#29

The agent page draws its delegation graph from the real trace tree, so you can follow one decision down three levels.

web#25

The dashboard is rebuilt on the prediction-market block: a rolling twenty-four hour token count, run and task-state pills that carry a word as well as a colour, fact chips and a cost footer, with an Agents tab in the sidebar.

web#32

The usage and billing pages stream their figures instead of polling for them.

web#30

Chat shows what an agent is doing between replies, and the thinking indicator only claims thinking when there is some.

web#26

You can delete an agent, behind a confirmation that names it.

web#33

A 404 page that types the path you asked for and refuses it, and an error page that shows the server’s real message.

web#28

Costs come from the billing catalogue’s real list prices, on the models list and on every usage figure derived from it.

platform-api#25

Search, the last admin domain still on fixtures, goes live.

admin#10

A run’s stream is a subscription that lives as long as the session, so the polling moved off your browser and onto the server.

platform-api#24

Model pricing in the wizard is honest: a model nobody serves is not free.

web#37

The bash tool is marked non-functional rather than promised, because the cluster it runs on does not permit the sandboxing it needs.

web#43

The transcript stops wiping itself every forty-five seconds, and a late live frame no longer produces a duplicate reply.

web#29

A dropped stream reconnects on its own, backing off between attempts instead of failing the run.

web#35

One broken agent card no longer freezes the whole dashboard grid.

web#27

The wizard’s list of built-in tools matches what the agent runtime actually loads.

web#36

The agent resource shown in the review step is the one that will be written.

web#41

Agent cards and the roster carry each agent’s own numbers, streamed per agent rather than shared across the grid.

web#34

The admin console’s clock runs, so live data stops being windowed out of view.

admin#9

Stats, usage, runs, quota and the tool catalogue go live, and the create wizard writes a real agent at the end of it.

web#11

The agent page has a live run loop: sessions, a composer that opens one with your first message, the streaming run and a real usage chart.

web#17

The agent page has a chat surface for talking to a running agent.

web#18

Runs stream over A2A, and a stream you dropped can be re-attached from where it left off.

platform-api#9

Your agents can delegate to subagents, declared as a tool of type Agent on the agent resource.

platform-api#16

The wizard offers live delegation candidates, so you pick subagents from the agents you actually have.

web#15

The wizard shows the platform tool library and the runtime’s built-in tools, marked as included.

web#13

The delegation graph is read from the trace tree the run left behind.

platform-api#8

The dashboard’s numbers arrive as a read and as a stream, so an open page keeps up on its own.

platform-api#4

Notifications are served by the platform rather than read from fixtures.

platform-api#7

Every tenant gets the agent baseline provisioned into its namespace, and the platform tool library merged into its catalogue.

platform-api#12

A sidebar shell over real routes, and a sign-in that carries the product rather than a plain form.

web#9

The admin console watches service health as a stream and reports measured uptime.

admin#6

The admin tenant and fleet views read real tenants and real agents across every namespace.

admin#8

Every figure carries a badge saying where it came from, at the component that renders it rather than once per section.

web#10

Tenant agents are fenced off the cluster’s private ranges, and cluster tools never reach a tenant catalogue.

platform-api#14

Every data surface resolves to something — a figure, an empty state or an error — so nothing sits on a skeleton forever.

web#12

Staleness on the agent page is a statement about how fresh a figure is, not about whether the agent is busy.

web#20

Creating a session stops claiming an idempotency the platform does not offer.

web#19