Plans, seats, pay-as-you-go and credit grants
Paid plans now deliver the entitlements the pricing page sells: plan credit grants land in expiring buckets on top of the universal daily allowance, and metered pay-as-you-go bills overage through Stripe Billing Meters, with Stripe Tax applied to every session that charges. Seats became a real axis — a seat is a person or an Agent — persisted, enforced and adjustable, with the allowance and the per-seat price shown on the Billing page. On a self-hosted install the purchase surfaces stay behind the payments flag until you configure a provider.
Learn more →Fleet: isolated task worktrees, node affinity, a signed Windows launcher
A node can now provision an isolated worktree per Task, so when a job arrives carrying a repository workspace, two runs on the same machine never share a checkout — a job without one still runs in the node working directory, and whether a Task gets its own workspace at all is the per-Work Task isolation setting. Per-Agent node affinity pins one Organization Agent to one specific box — the workstation with the right toolchain, the node that already holds the checkout — though it is API-only today, with no picker on the Agent page yet. Windows nodes gained a signed job launcher, and a task step can be allowed to use the credentials the node itself already holds instead of copying keys into the cloud.
Learn more →Invite a person into an Organization, end to end
Invitations are a real flow now: issue one from the members section in Settings, the recipient accepts it at /org-invite/<token>, and membership is created without anyone touching the database. Ownership guards still gate everything the Organization owns, and per-Work roles remain the fine-grained permission model.
Learn more →Teams can be staffed with people, not only Agents
A team roster now holds humans alongside Agents, with a manager Agent at the top, and the Org Chart draws both while the Agents Chart keeps to reporting lines. Membership stays organizational rather than a permission: access to a Work is still granted per Work through Work Members.
Learn more →Navigation consolidation
Three orphan destinations found their homes. Teams became a hub with the org chart and the agents chart beneath it, Meetings moved into Memory next to the review queue and Files, and Skills moved under Agents — where they are actually bound.
Environments — named runtime sandboxes, assignable per Agent
Settings → Environments defines the sandbox an Agent runs inside: pip and npm packages, Unrestricted or Limited networking with an allowed-host list, and an "Allow package managers" toggle, saved as a draft or published. Assign one to an Agent and every run resolves through the published row; an Environment an Agent still references refuses to be deleted, so an assignment can never dangle.
Learn more →The Agent Workbench — Capabilities, Collaborators, session timelines
The Agent detail page grew into a ten-tab workbench. Capabilities holds the per-tool grant matrix, skill bindings and an init script behind one composed read; Collaborators is an explicit allow-list of the sub-agents this Agent may delegate to; and a session opened from /agents/sessions/<runId> replays the whole run — tool calls with redacted argument and result previews, and every file it touched.
Learn more →MCP client connections
Register an external MCP server once under Settings → Connections, probe it with the built-in test, then bind it to the Agents allowed to use it. Its tools appear in those runs alongside the built-in ones, filtered by the same tool-grant matrix and the same permission flags — which is how an Agent reaches a capability Ever Works does not ship itself.
Learn more →Repositories registry
Settings → Repositories is the account-level list of Git repositories that are not attached to a Work — the service repo an Agent has to read, the design-system repo it copies tokens from, the private tooling repo it runs a script out of. Add one by hand or import it in a click from a GitHub App installation, give it a mount directory, a credential pointer and up to eight masked seed .env files, then attach it to an Agent the way you attach a Skill. Today an attachment is a durable, audited grant rather than a checkout: the registry, the GitHub App import and the per-Agent binding are live, but no run executor mounts an attached repository yet — a Task provision spec carries the list only as advice, and multi-repo mounting is still coming.
Learn more →Goals: Definition of Done, budgets and an orchestrator loop
A Goal watches one number, read from a metrics provider on a schedule. August added the autonomy layer around it — a Definition of Done, budgets, and an orchestrator loop that advances the Goal on its own. That loop is new: it has no end-to-end browser coverage yet, and it stops short if the Goal has no Agents to draw on.
Learn more →Inbox — the operator message center
One list for everything the AI workforce needs a human for: a question from an Agent that parked its own run, a pending proposal, an escalation, a budget notice. Pick the recommended answer or write your own, and the reply routes back as steer, resume, approve, reject or resolve — with an unread badge in the sidebar so nothing waits unseen.
Learn more →Task schedule modes, workflow templates and sub-tasks
A Task now picks Run once, Scheduled or Recurring, and refuses to hold a one-shot slot and a recurrence at the same time. The New Task form gained a From template mode that expands a workflow template into a parent Task plus one sub-task per step in a single transaction — dependency blockers and approval gates included — and the Sub-tasks checklist on Task detail counts how many children are done.
Learn more →Triggers get a home, and a second source: platform events
Signed inbound triggers arrived in July (below). August gave them a real surface — a Tasks → Triggers tab with a detail page per trigger — and a second way to fire one: an event matcher on the ingest spine, so an ingested platform event whose source, kind or Work matches your rule spawns the Task instead of an outside caller. A trigger can now build its Task from a template as well as from instructions, hide it from the board, be rehearsed with Test fire, and be audited through a recent-fires log you can read when an external system swears it called you.
Learn more →Memory Files — one file browser, with Git sync
Memory gained a Files panel spanning both upload paths — chat attachments and Knowledge Base originals — organized into folders you define, with preview, move and download on every row. Point a folder at a repository and Sync now commits its subtree to GitHub, reporting how many files were committed and how many were skipped. Unfiling a file and deleting a folder never destroy bytes.
Learn more →Skill invocation slugs and companion files
A Skill can carry a slash command of its own — type /plan in the composer and that Skill comes in — and ship companion files alongside its SKILL.md. Slugs are unique per account, so a collision is refused by name instead of silently shadowing the Skill you meant.
Learn more →Saved workflows run as dispatched jobs
Workflow graphs are persisted rows with a CRUD API behind them, and a saved graph can be executed as a dispatched job that leaves a run record you can open afterwards. A delegated child run is confined to the scope it was admitted under, and delegation depth is derived server-side rather than trusted from the caller.
Learn more →Per-Work runtime environment variables
A deployed Work can carry its own allow-listed runtime environment variables — the payment keys a storefront needs — stored server-side, listed with masked values on the Deploy tab, and delivered by both the Kubernetes and the GitHub path. The allow-list is the point: a Work cannot be handed arbitrary secrets.
Learn more →A model picker in the chat composer
Pin a model beside the composer: the provider default, one of the provider's configured simple / medium / complex tiers, or anything in its catalog. The pin is stored on the conversation row, so it survives a reload on another device. Voice provider selection moved into Settings at the same time.
Learn more →