Human in the Loop

Human-in-the-Loop Approval Flows in Ubik Studio

Human-in-the-Loop (HITL) is a foundational principle in Ubik Studio that ensures AI agents never act unilaterally. Rather than executing actions automatically, agents show their intent before proceeding—they propose research queries, suggest new documents, recommend edits, or request clarification, then wait for your approval. This approach preserves human agency and control while enabling the speed and efficiency of automated workflows.

In Ubik Studio, HITL operates through a system of approval cards. These cards appear in your workspace whenever an agent needs your decision. You can approve the proposed action as-is, reject it entirely, or edit parameters before approving. This flexibility means you maintain oversight without micromanaging every step. The agent proposes; you decide. Together, this creates a controlled, transparent, and auditable workflow.

Configuring AI Autonomy Levels

Ubik lets you tune how much control you retain across different workflows. You can choose from several scales depending on the task:

• Full Spectrum (Manual → Autonomous): A five-level scale from fully human to fully AI. Suitable for project-level or org-level generation policies. Options range from Manual through Assisted, Collaborative, and Directed to Autonomous.
• Writing Agent (3-level subset): Framework, Guided, or Full AI—maps to the orange→green range. Used in the writing-agent create and edit widgets.
• Code Review (custom domain scale): Domain-specific levels built from the shared palette—e.g. Review only, Suggest fixes, or Auto-fix. Any number of stops works.
• Binary (On / Off): A two-stop scale for simple "human vs AI" decisions. Works well for per-step toggles in a pipeline (Manual vs Automated).

These scales let you dial in the right balance of oversight and automation for each workflow.

How Approval Cards Work

Approval cards are the user interface for HITL. Each card displays the agent's proposed action with all relevant details: what the agent wants to do, what parameters it will use, and what it expects to happen next.

• Title and Description: Clearly states the action type (e.g. "Run Research," "Run Create Workflow")
• Parameters: Shows the specific inputs the agent will use—search queries, document names, editing instructions, or clarification options
• Action Buttons: Approve to proceed, Reject to decline, or Edit to modify parameters before approving

The edit option is particularly powerful. You can change a search query to be more specific, rename a document before creation, or adjust the scope of an edit without stopping the workflow. Once you've made changes, you approve the edited version, and the agent executes with your modifications.

Research Approvals

Research actions require approval before the agent queries external sources or processes information. Ubik Studio shows several types of research approval cards.

Initial Research

The Run Research card shows:

• Initial research task — A high-level topic or query (e.g. "Efficient Attention Mechani...")
• Detailed research prompt — The full instructions the agent will use, such as: "Find 5–10 key papers on how attention mechanisms in transformers scale with sequence length and what efficient alternatives (sparse, linear, sub-quadratic) have emerged since 2020. Focus on empirically validated approaches that appear in production systems."
• Research mode — Quick Find, Literature Review, or Deep Dive, plus a More dropdown for additional options
• Other settings — Summary of constraints (e.g. "Academic only," "5 min" estimated time)

You can approve to let the agent search, or edit the task, prompt, or mode to refine the scope. For example, you might switch from Quick Find to Literature Review for a more thorough search.

Followup Research

When an agent wants to continue an existing research session, it surfaces a followup card showing:

• The original research context (what was already found)
• The followup query or action the agent proposes
• The rationale for the followup

This allows you to confirm that the agent's next step makes sense or suggest a different direction.

Read URL

For single URL lookups, the Run Read URL card shows:

• Title — A short label for the task (e.g. "NY DFS Crypto Regulations")
• Starting URL — The exact URL the agent will access (e.g. https://www.dfs.ny.gov/industry_guidance/industry_letters)
• Task — What the agent should do, e.g. "Check for any recent announcements, industry letters, or proposed regulations related to cryptocurrency or virtual currency from the past 3 months."
• Navigate — Options for how the agent will proceed
• Other settings — Configurable timeout or duration (e.g. "5 min")

You can approve to proceed, use Approve w/ edits to tweak the URL or task, or reject if the source is unreliable or off-topic.

Writing Approvals

Writing actions in Ubik Studio require approval before the agent creates new documents or modifies existing ones.

Create Document

The Run Create Workflow card shows:

• Title — The proposed document title (e.g. "Benchmark Contamination: k...")
• Priority — Low, Medium, or High, so you can signal urgency or importance
• Task description — Detailed instructions, e.g. "Create a 500-word summary document synthesizing key findings and usable quotes from the three benchmark contamination papers. Structure as: (1) Executive summary of the contamination problem, (2) Key findings on data leakage and preference leakage severity, (3)..." with more structure as needed
• Notes to use — Tag-like chips showing which workspace notes the agent will draw from (e.g. "key-point: N-gra...", "evidence: Stand: +8 more"). An Adjust button lets you add or remove notes before approving
• Run settings summary — Parameters such as "500 words," "4 required sections," and page or formatting constraints

You can approve to let the agent create the document, reject if the topic or scope isn't right, or edit the title, task, priority, or notes before approving.

Edit Document

The edit card shows:

• The document being edited
• A summary of the proposed changes (e.g. "Add three paragraphs to the Methodology section" or "Revise the Conclusion based on new findings")
• The sections or content areas affected
• The rationale for the edit

You can approve the edit as proposed, reject it, or edit the instructions before approving—for example, you might narrow the scope to a single section or expand it to include additional context.

Questions and Answers

Agents sometimes need clarification to proceed effectively. Ubik Studio uses QA cards to request information from you in a structured way.

Questions

A questions card shows:

• The question being asked (e.g. "Should this document focus on technical implementation or business value?")
• Response options, which can be:
  • Multiple choice: A set of predefined options you select from (e.g. "Yes / No / Unsure")
  • Freeform text: An open-ended field where you type a custom answer
  • Hybrid: Both options, allowing you to choose a preset or provide your own

The card displays the agent's reasoning for asking, so you understand the context.

Answers

After you respond, an answers card appears in the chat summarising:

• Your response (the choice you made or text you entered)
• The timestamp of your answer
• How the agent will use your answer (e.g. "This will inform the document's target audience section")

These cards create a permanent record of all human decisions, making it easy to trace how the workflow progressed and what decisions shaped the outcome.

Why This Matters

Human-in-the-Loop approval flows are essential for three reasons: control, clarity, and quality.

Control: You remain the decision-maker. The agent proposes; you decide. This prevents unintended actions and ensures the workflow aligns with your intentions. If the agent misunderstands your goal, the approval checkpoint catches it before wasted effort occurs.

Clarity: Approval cards make the agent's reasoning visible. You see exactly what the agent plans to do, why, and with what parameters. This transparency builds trust and helps you spot misunderstandings early.

Quality: The ability to edit parameters before approving means you can refine the agent's work without restarting. A slightly off search query, a document title that needs adjustment, or an edit scope that's too broad—all can be corrected in a single approval step, keeping the workflow moving while maintaining quality.

Together, these approval flows enable controlled autonomous work: agents handle the heavy lifting of research, writing, and iteration, while you maintain oversight through lightweight, high-value checkpoints. The result is faster, more reliable workflows that respect your judgment and expertise.