QuoxCORE 2: Your AI ops, on rails you can see

The 11 things shipped

1. Sprint-1B + Crystals foundation

The base layer for everything else. We introduced the Crystal primitive, a reusable, governed skill that emerges from actual task executions rather than manual authoring. This phase wired the Crystal library, the promotion UI, and the execution history that makes task-time crystallisation possible. If you have been saving the same prompt over and over, Crystals replace that pattern with something that audits itself.

2. Execution plumbing

The collector service now streams execution events in real time to the workflow canvas. Nodes light up as they run. Tool calls appear in a side panel as they fire. You can watch a 30-step workflow execute without refreshing the page. This is the plumbing that makes the live canvas overlay in later phases possible.

3. Agent taxonomy clarification

We clarified the Assistants vs Agents distinction that was causing confusion in the builder. Assistants are supervisors: they hold conversation context and route tasks. Agents are workers: they execute discrete tasks with tools. The mode discriminator (agent:standard, agent:agentic, agent:browser) is now the runtime switch that prevents a looping agent from being invoked where a single-response specialist is expected.

4. Tools Deep: four install paths, one governance model

The Tools view got a complete rebuild across seven sub-phases. You can now install tools via OpenAPI spec, MCP server, Git repo URL, or Python script. Each install path produces tools that are governed the same way: budget-gated, AOCL-wrapped, and assigned to specific agents via the matrix.

The Tool Audit tab shows a coverage map: which tools are assigned to which agents and which have no assignments at all. Install-time budget constraints prevent a single tool from consuming the execution budget in one call.

5. Workflows Unified

The workflow builder, library, versioning, and live execution overlay all ship as one unified surface under a tabbed hub. The builder canvas now shows a live overlay during execution: nodes pulse, go green on success, go red on error. Workflow versioning captures every save as a restorable version with a node-level diff.

The library has a Curated tab with 12 Quox-authored templates and a My Library tab for user-authored workflows. The ?fromObjective= URL parameter pre-populates the builder from an existing objective.

6. Observability + Mission Control

Mission Control is the single-screen view of everything your agents are doing: active objectives, live runs, pending approvals, budget burn, and the evidence chain. The schedule bridge now surfaces workflow trigger:schedule nodes in the /schedules tab so cron-triggered workflows are visible alongside manually triggered ones. The QuoxFlow evidence and observations proxy means AOCL traces from workflow executions are readable directly in the Mission Control waterfall without leaving the dashboard.

7. Studio cleanup + User-owned Prompts Library

We removed surfaces that existed to fill space: AgentAutonomy and VideoStudio views are gone. ImageGallery moved to Work. The Prompts Library is now genuinely user-owned: prompts are stored server-side, scoped to your org, and persist across browser clears. We deleted the 75 scraped third-party prompt examples that were sitting in the library without attribution. If you built any prompts in that library, they are now the only prompts there.

8. Governance and reliability

The largest phase. Six sub-phases, each targeting a specific enterprise gap:

• Triggers UI (P8.A): a dedicated /triggers view for managing scheduled and event-driven workflow triggers.
• Cost budgets (P8.B): four-scope budget enforcement (org, objective, execution, iteration) with a hard-stop dialog when limits are breached, and a Costs tab in the builder showing per-node spend.
• Approval policies (P8.C): policy templates, ghost-render gates in the builder, and inbox policy badges so approvers see which policy triggered the request.
• Reliability (P8.D): real concurrency via util:wait abort, multi-stakeholder threshold for HITL nodes, and loop-item subworkflows so each iteration runs in isolation.
• Engine Validation Suite (P8.E): five nightly test batches covering nine surfaces. Any regression in execution, budgets, policies, evidence chain, or triggers fails the suite before it reaches production.
• HITL org routing (P8.F): workflow inbox items now land in the correct org. Previously, executions without an explicit owner defaulted to the system org. Fixed.

The governance phase was triggered by a real incident: a 68-hour stuck execution that required a direct database kill. P8.D, P8.E, and the Engine Validation Suite are the response.

9. Task-time Skill Loop: Crystals that form mid-run

Before a task loop starts, Quox Brain retrieves relevant Crystals from the library and injects them as context. Mid-loop, it observes the agent's actions and drafts a Crystal candidate based on what is working. At the end of the task, the candidate is presented with a diff and a promotion button. If you approve, the Crystal is canonical and available to all future runs.

If you reject, the draft is discarded. No Crystal reaches the library without explicit user approval.

10. Tie-it-together UX: the Home view as ops dashboard

The Home view is now a proper Mission Control, not a chat interface with a sidebar. The Goal Lane shows active objectives with progress and budget at a glance. Quox Brain fires proactive suggestions when it detects a pattern worth crystallising or a workflow worth running: six starter rules ship out of the box.

The "Save as Objective", "Save as Workflow", and "Save as Crystal" CTAs let you promote any conversation to a persistent, governed artefact in one click, with the builder pre-populated. Live canvas lighting and Run-now Crystals mean the home view is where work actually happens, not just where it is planned.

11. Engine Validation Suite

Five nightly test batches covering execution correctness, budget enforcement, policy gates, evidence chain integrity, and trigger reliability. The suite runs against the live QuoxFlow engine, not mocks. Any regression in these nine surfaces fails the build before it reaches staging.

What each tier unlocks

QuoxCORE remains free. The free tier includes the workflow builder, 49 agents, the Crystal library, Mission Control, the Prompts Library, and the full AOCL/VOLT/WARD evidence chain. You can see everything your agents do. You can build and run workflows. You can version and restore them.

Compliance Suite ($499/yr) Task-time Crystals (mid-run drafting), Quox Brain v1 (proactive suggestions, 6 starter rules), per-step cost budgets with hard-stop enforcement (four scopes), SOC2/HIPAA/GDPR/ISO audit export, priority support.

Developer Kit ($199 one-time) Approval policies and policy templates, reliability tier (real concurrency, loop-item subworkflows, multi-stakeholder HITL), AOCL audit export, Layer SDK, Layer Playground, custom AOCL layer plugins.

What is coming next

• BQ-Agentic-Migration: Browser-Q gets a proper agentic execution loop. The observation fix that proved the concept ships first; the full migration follows.
• Crystal Marketplace: Browse and install Crystals authored by the community. Rating, versioning, and VOLT provenance trace included.
• Quox Brain v2: LLM-powered suggestion engine. v1 ships six handwritten rules; v2 replaces them with a trained classifier that surfaces suggestions based on actual usage patterns, not hardcoded heuristics.
• P9.5 Critic-gate: A mid-run validation step where a second model reviews the Crystal candidate before proposing promotion. Reduces false positives in the task-time crystallisation path. Deferred pending compute budget review.