What changed on VegaDūta

Workflow expression language: node fields can read the trigger payload and earlier nodes' outputs, with a live preview and a lint panel. Guide at /help/workflow-expressions.

Workflow data pinning: save a node's output on the draft so later steps can be built and tested without re-running it. Draft runs use pins and are labelled as such; live runs (webhooks, schedules, channels, share links) never do.

Team / Swarm: a lead agent delegates to two or more member agents in parallel. Depth 1 only, at most 8 branches per call, each branch stopped after 240 seconds. Three presets (research crew, review pair, support triage).

Memory page: read, edit and delete what your assistants remember about you, including per-project notes.

Eval reliability: pass^k promotion gate (a case must pass on every repeat), confidence intervals on scores, and paired comparison between runs.

Multi-agent delegation now runs under a shared per-run budget by default (2 US dollars, an estimate used as a runaway guard). A run that uses it up stops with a stated reason.

Eval datasets, reliability statistics and Agent Lab knockout and consistency runs are open to Free workspaces with smaller run limits.

Public agent directory at /directory for agents a workspace has chosen to publish.

Agent Lab in the playground: a run graph of each turn (plan, tool calls, guardrail decisions, handoffs, reply), fork-from-a-message with side-by-side comparison, prompt knockout, and a consistency check. API under /api/agents/{id}/lab.

Agent Skills: SKILL.md packages that agents load only when a request matches. 14 built-in skills, a Skills page with .zip import and export, and agent-drafted skills that stay switched off until a workspace admin enables them.

Staging and production labels on agent and workflow versions, with restore. Promotion to production can require a passing eval run.

Simulation testing: bounded multi-turn conversations between an agent and built-in or custom personas, each ending with a stated reason.

Workflow nodes: bulk email (recipient cap, send-rate limit, unsubscribe and suppression list), CRM lead (writes to a built-in leads list with stages), and voice menu (keypad options on a call). The node catalog is now 21 types.

Email reply and bounce detection: a reply to outreach is recorded on the lead and moves it to the replied stage.

App Studio: Figma frame import, GitHub repository import and sync, a bounded self-repair loop on build errors, a database role per project, end-user sign-in for published apps, and disposable apps that archive themselves. Custom domains can be verified but do not serve apps yet.

Background workflow runs: start a run and poll for its status instead of holding a request open.

Dynamic MCP STDIO servers (npx/uvx packages a tenant admin registers) now spawn inside an isolated mcp-stdio-runner container - no platform secrets, no route to the data tier, non-root, all capabilities dropped, memory and PID limits, and a custom seccomp profile that denies unprivileged user-namespace creation, ptrace and the kernel-attack surface. Previously the package ran as a child of the core JVM holding APP_ENCRYPTION_KEY.

Outcome-based pricing substrate: billable resolution events with evidence classification, idempotent dedupe keys, per-event price snapshots, dispute handling, and a tenant-facing summary API (GET/POST /api/tenant/resolutions).

Public status page at /status backed by a new unauthenticated GET /api/public/status endpoint reporting live component health.

Public API reference at /api-reference covering authentication and the platform's REST surface.

Agent eval gate: dataset/case/run/judge infrastructure with a per-agent quality gate that can block promotion on eval verdicts (V260).

Workflow run pause/resume and input requests (V261, V271) so long-running flows can wait on a human answer.

Community template marketplace with star ratings (V251, V264) and a public gallery at /templates.

First-party plugin catalog (IDE/browser extensions) at /plugins, served by a public GET /api/plugins/catalog.

Public share links for agents and workflows (/a/<id>, /w/<id>) backed by /api/share/** - read-only snapshots that never leak edit surfaces.

Tenant SSO / SCIM provisioning (V253) for enterprise identity rollouts.

Billing webhook event ledger (V250): provider webhooks are deduplicated on a durable unique key so a redelivery can never double-apply a money event.

Commerce product feed (V252) and public /api/commerce/feed/** endpoints for storefront-style agent catalogs.

Human handoff for channel sessions (V214): an operator can take a live WhatsApp/Telegram/etc. conversation away from the agent (AGENT / AWAITING_HUMAN / HUMAN states) and release it back.

Public documentation hubs: /docs plus the /integrations, /ai, /platform and /tutorials libraries, all server-rendered and crawlable.

Turn-level feedback signals on agent replies: thumbs, star rating, edit, and resolved/unresolved.