Governed Autonomy
SQL and database connectors, an AI report swarm grounded in your real data, end-to-end resource access control, self-governing model routing, and an OAuth provider for external assistants.
What’s new
v0.4.0 made the work verifiable. Aimless v0.5.0 makes it connected and governed: the AI now reaches into your databases, writes full reports from real data, and runs every resource — agents, sources, models, and budgets — inside explicit access and spend boundaries. More happens on its own, and all of it happens inside the lines you draw.
Connect a database, ask in plain language
Connect Postgres, MySQL, or any supported SQL source and the orchestrator routes data questions straight to it. It writes the query, self-corrects against the real schema when a query fails, and binds the result into dashboards — no hand-written SQL. A guided connection wizard reads from a dialect registry so each engine’s quirks are handled for you, and connected databases are now discoverable by agents as first-class sources alongside APIs.
AI reports, grounded in your data
Ask for a report and a dedicated report swarm assembles it — pulling from your connected apps, live widget data, and SQL sources rather than from thin air. The finished report renders inline in chat with a download button, persists to a permanent /reports/<id> URL, and is archived to your organization’s document vault and object store so it’s searchable later. For autonomous runs, the agent now writes a human-readable summary to your inbox and asks for approval before it acts.
Resource access control
Every resource is now access-controlled end to end. Data sources are organization-scoped, a unified authorization chokepoint governs each request through a shared grants model, and you can grant or restrict an individual agent’s access to a specific source — with inheritance by default and explicit overrides where you need them. A new Access panel gives you one place to see and manage who can reach what. API-key connectors are visible and callable by agents under the same rules.
Self-governing model routing
The model router now governs itself. When a provider delists or decommissions a model, a catalog reconciliation pass detects it and automatically routes around the dead model instead of failing requests. Registry read failures are distinguished from a genuinely empty model band, and a reliable free-tier fallback keeps low-cost requests flowing even when a specific free model disappears.
Spend guardrails
Every LLM spend path now respects a per-organization monthly hard cap by default, so a runaway automation can’t quietly burn budget. Every executable automation — scheduled tasks, polls, and swarm crons — is surfaced in Settings → Automations where you can see and pause it, with the same organization guard applied on the polling path so there are no hidden runnable jobs.
Marketing automation on a schedule
The marketing swarm now runs on its own cadence: a weekly SEO audit and a blog-post pipeline execute on cron, a trend-researcher phase pulls live web search to ground topic planning, and the whole pipeline publishes and crawls from cloud workers in production. Every marketing cron shows up as a visible, pausable task in your Overview.
Open to external assistants
Aimless is now an OAuth 2.0 provider. External assistants — including Microsoft Copilot Studio — can connect through a standard authorization flow with a consent screen, bearer authentication, an external JSON chat surface, and a published OpenAPI spec. The authorize flow resumes cleanly after sign-in, so connecting an outside client just works.
Reliable media and a clearer reasoning view
Generated images and media now render exactly once, reliably, with a live placeholder while they’re still generating — no duplicate renders, no missing assets. The Glass Box reasoning spine groups a swarm’s phases into a single phase card instead of scattering them, and the inbox threads conversations per agent with plain-language run digests and clickable swarm pills on each task.
Streaming and agent reliability
Replies stream token by token again — the output filters no longer buffer the whole message before showing it. When an agent describes a tool call in prose instead of actually invoking it, the stream now detects and executes it for real. Retries no longer double-execute or double-bill, the cache-reuse prompt stopped nagging on every endpoint (freshness auto-resolves, decisions stick per run), and the skill picker is now context-aware — it won’t pull in an external skill that duplicates a native tool.
Infrastructure improvements
- Multi-tenant hardening: notification producers are organization-scoped to prevent cross-tenant leakage, and waitlist tooling is scoped to the requesting organization.
- Fixed an invited-user login loop by enabling Google account linking, plus a one-time beta credit for allowlisted signups.
- Dashboard binding correctness: self-binding charts render live data through the registry path with numeric-string coercion and role-based field mapping, instead of falling back to demo content.
- The database bind pipeline is hardened with idempotent sources, foreign-key upserts, an internal-key profiler and poller, and automatic saved-query slug de-duplication on collision.
- Embeddings call the REST API directly, dropping the OpenAI SDK dependency.
- Resource pages gain a marketplace-style filter toolbar, a compact list view, add-new buttons, and an explicit-connect vault flow.