We’re building Ever Works in public. Some things may be incomplete, missing, or broken while we continue improving the platform.We’re building Ever Works in public — expect a few rough edges.
Control & safety · Human in the loop
An autonomous workforce is only useful if you can say how autonomous. Ever Works splits that into layers you set separately — what an Agent may do alone, which tools it can reach, what it has to ask about first, and what it may spend — and each layer is a setting you attach to an Agent, a Work, a Mission, an organization, or your whole account. None of it is hardcoded, and every default is the careful one.
Six independent mechanisms. Run every one of them at its default and you have a workforce that cannot surprise you — then loosen exactly the one that is slowing you down.
Each Agent has a dispatch mode on its Guardrails card: Require approval queues every side-effectful action, Autonomous lets unflagged ones through. Narrow autonomy to an auto-approve list of action types, and name the types this Agent may never take at all — blocked applies in both modes.
Pending proposals appear in the Action approvals block on your dashboard home, each showing what was proposed and any risk flags — destructive, cross-scope, high fan-out, budget override. Approve or reject per row, or clear the whole visible queue with Approve all.
When a run gives up — a merge the policy refused, a check loop that used every attempt, a budget that stopped the work — it writes an escalation scored for how likely it is that a person genuinely has to act. The queue orders by that score before recency, and the items arrive in your Inbox.
Before a run starts it passes an admission gate: concurrency per Work and per organization, your plan’s session ceiling, and a credits pre-check. A run that does not get through is parked with its reason attached — concurrency-limit, or insufficient-credits — and stays visible in the Sessions view instead of vanishing.
Cap spend on your account, a Mission, an Idea, a Work, and each Agent. An Agent budget runs hourly, daily, weekly, monthly, or unlimited; make it hard and the next AI call is refused before it runs, or soft and you are alerted while the work carries on.
Pause after failures — three consecutive failed runs by default, adjustable from one to twenty — moves an Agent to paused instead of retrying forever. The doom-loop detector spots a run cycling on the same failure with no progress, stops it, and escalates with the evidence rather than paying for a fifth identical attempt.
Guardrails decide what an Agent may do. These decide what it can reach in the first place.
Create agents, assign tasks, edit skills, edit instructions, spend budget, commit to repo, open pull requests, call external tools. Every flag is off on a new Agent, and a flag that is off means the tool is never assembled for the run — the model never sees it and cannot ask for it.
Allow or deny individual tools on the Agent’s Capabilities tab. Grants resolve down four scopes — tenant, organization, Work, Agent — and each step may only narrow the step above it, so a deny is permanent and nothing further down can re-enable it.
A Skill can declare the tools it exists to drive. When the matrix denies all of them the Skill is suppressed instead of injected, so an Agent is never handed instructions for a capability you deliberately took away. Keep one of the tools and the Skill stays active.
Sub-agent delegation is opt-in per pair: until you enable a collaborator on the Collaborators tab, an Agent may only delegate to itself. Delegation depth and per-run fan-out are both capped, and a child’s scope is always the intersection with its parent’s — privilege only shrinks going down.
Assign a published Environment and every run of that Agent starts from the same toolchain and the same egress rules: Unrestricted, or Limited to the hosts you list. See Environments for how one is written, published and assigned.
Merge policy decides whether an Agent may land its own pull request, resolved at tenant, organization, Work or Agent level. Quality gates decide whether a pull request opens at all — a red build, test, lint or type check opens nothing.
From an Agent deciding it wants to do something, to you deciding whether it may.
A side-effectful action — spawn a sub-agent, schedule a task, send a message, override a budget — is written down as a proposal rather than simply taken.
A deterministic scorer flags destructive, cross-scope, high fan-out and budget-override actions, then the Agent’s guardrails are evaluated first match wins. A blocked type is rejected and recorded, never silently dropped.
An unflagged action on an autonomous Agent proceeds and is recorded as auto-approved. Anything flagged queues for a human whatever the mode says, and the dispatch gate can still park the run for capacity or credits.
Approve or reject in the dashboard queue, or answer the same item in your Inbox. A decision is final — re-deciding is refused, not silently flipped — and the outcome lands in the activity log and, where it touched content or code, as a commit in your own Git.
Every default here assumes you have not watched this Agent work yet. Autonomy is granted deliberately, one layer at a time, once the run history gives you a reason to grant it — and each layer has a reset that puts the strict setting back in one click.
Agent questions, approval requests, escalations and notices land in one Inbox — and reach you in-app or on your own chat channels, tuned per event.
Learn more →Who may merge, when, and into which branches — a setting you resolve at tenant, organization, Work, and Agent level.
Learn more →Declare the checks a change must pass. The platform runs them for real — and a red check opens no pull request.
Learn more →Every run is metered in credits, recorded in a ledger you can read — with daily free credits, plan entitlements, and free runs on your own keys.
Learn more →Run many agents at once from one cockpit — live status, current activity, tokens and cost — and steer, interrupt, or resume any of them.
Learn more →Named AI employees — a CEO, a CTO, a Researcher — that run your Missions, Ideas, and Works around the clock.
Learn more →Named runtime sandboxes you assign per Agent, credentials resolved from your own secret store, and a registry of the external MCP servers and repositories your agents may reach.
Learn more →Describe what you want to build. Ever Works researches it, ships it, and keeps it improving — content and code, owned in your own Git.