Earned Trust
A system that grades its own honesty, one unified credential vault, sharing that takes a yes, visible model swaps with savings, and continuous deployment.
What’s new
v0.6.0 finished the work. Aimless v0.7.0 is about trusting it. The system now measures how often its word matches its work, keeps every credential in one vault that proves a connection is real, asks before it assumes, and shows you every model swap and what it saved. Sharing takes the other person’s yes. Autonomy you cannot inspect is just a guess; this release is about making it earned.
A system that grades its own honesty
v0.4.0 verified the data. v0.7.0 verifies the claims. Every run is now checked for false success: cases where an agent reported finishing something it could not actually prove. That rate is tracked over time and surfaced on a trust card in your Overview, so you get a number for how often the system’s word matches its work, not a feeling. Agents are prompted to verify before they claim and to say plainly when they could not, and full run traces export end to end so any decision can be inspected after the fact.
One vault for every credential
All your provider credentials now live in a single unified vault, promoted to the front of the Vaults page instead of being scattered across per-connector screens. Each credential carries a verified-versus-claimed status, so you can tell a connection that actually works from one that only looks configured. Saving a custom connector is reliable, and a live badge shows real traffic running against it.
Sharing takes a yes
Private-by-default now has a complete handshake. When you share an agent, report, dashboard, or data source, the recipient is notified and the resource stays private until they accept. Accepting creates an explicit grant; declining leaves it closed; and re-sharing re-offers cleanly. Access is something both sides agree to, never something that quietly happens to you.
Ask before assuming
A vague request no longer becomes a confident guess. The orchestrator asks one focused clarifying question, then builds. And when a dashboard’s underlying data is too thin to mean anything (all zeros, or a single entity), a pause-and-ask gate stops to check with you instead of rendering an empty-looking result. Quality over speed, by default.
Every model swap, in the open
Cost-aware routing is now fully visible. The reasoning timeline shows each model swap as it happens, names the model chosen and the one it replaced, and reports the savings as a percentage. Delegated agents show the model they actually ran on, so you can watch the whole routing policy play out across a multi-agent run rather than taking it on faith.
Sharper, safer communication
Agents now lead with the outcome instead of narrating process, ask before doing anything hard to undo, and close with a single clear message rather than trailing tool noise. A resumed conversation renders in full, with every decision, delegation, and code block intact, instead of collapsing into flat text.
Rolling out now
A few of these are landing as we publish this release:
- Always-on model-swap rows that show the savings on every routing decision, plus per-delegated-agent model attribution in the timeline.
- The pause-and-ask data gate that halts a low-signal dashboard build before it renders.
- Custom-connector credential saving hardening in the vault, with a clearer connection badge and a live traffic indicator.
- Mid-chain delegation recovery, so a sub-agent that narrates a result without returning it is caught and re-driven instead of stalling the run.
Infrastructure improvements
- Production web now deploys automatically on merge, closing the gap between “merged” and “live.”
- KPI cards reject non-numeric bindings at the source, so a categorical column can no longer masquerade as a metric.
- The system prompt now teaches outcome-first communication, action-safety on irreversible steps, and a single final message, applied across every agent rather than one at a time.