Standing Orders
Loops that run on their own toward a goal, another no-fake-data hardening pass, budget caps you can see with a live model catalog, and a Glass Box that shows the whole crew.
What’s new
v0.7.0 made the work inspectable. Aimless v0.8.0 lets it run unattended. Automations become standing Loops that keep working toward a goal instead of firing once. The no-fake-data stance gets stricter again. You can see and cap what a run spends. And the reasoning view now shows the whole crew, live, however long the work takes.
Loops: automation that runs on its own
Tasks ran once. Loops keep working. A Loop pairs a goal with a condition to watch, runs on a human-readable cadence, and can coordinate several agents toward several goals at once. Before a Loop ever runs, a preflight gate checks readiness: it confirms the right capabilities are connected, selects agents by what they can actually do, and provisions any missing connector the moment you approve it. Tasks and Loops now live in one unified work list with a guided setup overlay, so a standing operation is something you configure once and trust to keep going.
Still no fake data
The stance from v0.4.0 and v0.7.0 got stricter across every build path. A dashboard now stops hard when a connected source’s live read fails, instead of quietly falling back to mock numbers. Diagnostic and enum-discovery probe queries can no longer be mistaken for real data and bound to a widget. Widgets showing sample data wear a visible badge, so you are never guessing whether a number is live. Connector failures surface verbatim instead of being papered over with a confident-sounding excuse, and an edit that fails reports an honest failure rather than a misleading one. The data-integrity verifier is now a blocking gate, and a growing suite of evals guards every one of these paths against silent fallback to mock data.
Costs you can cap
Spend stopped being a mystery. Org budgets are ledger-accurate with a friendly cost-cap experience and a live wallet on the run bar, so you always know what a run is costing as it happens. When a single model fails or is overloaded, the system degrades gracefully down a fallback chain instead of failing the whole run. And model labels and prices are driven by a live catalog that refreshes automatically and decommissions stale models within 48 hours, so what you see quoted is what you actually pay.
A Glass Box that shows the whole crew
The reasoning timeline holds up the longer the work runs. Swarms render as a named crew of agents rather than anonymous phases, and an asynchronous report shows its crew relaying work right in the in-flight card. Long delegations tick a live elapsed timer so nothing ever looks frozen, repeated tool calls and internal discovery noise collapse instead of flooding the view, and delegated conversations render full-width with their tool activity inline.
Infrastructure improvements
- The data pipeline now has a single owner: the data analyst holds every binding step and reads connected REST and SQL vaults alike, so a connected-database dashboard builds deterministically and the orchestrator delegates to it automatically.
- Nested delegations keep their full tool set instead of dropping it mid-chain, fixing a self-delegation loop that could stall a multi-agent run.
- Database-backed builds fail fast on a slow or unreachable backend, deduplicate their bindings so a rebuild can’t exhaust quota, and batch their bind operations to collapse many sequential round-trips into far fewer.
- External assistant sessions (such as Copilot) now resolve to the correct organization, removing a dependency on a fragile per-user pin.
- The inbox unifies its unread badge and panel count behind one shared model and redesigns thread rendering.