How should users and AI agents share control of a task?

Let users co-plan and assign steps, interleave planning with doing, and hand control back on risky steps. Early studies show this steers agents without making them hard to use.

EvidenceEarly signal, treat it as a hypothesisLLM fitTested in lab or early LLM studies, likely carries over

Two builder-run system studies with small samples and simulated users.

system evaluation2 of 2 tested with LLMsEvidence from 2025–2026How we rate

Design it like this

Make the plan editable, let steps be assigned to the user or agent, and support easy handover.

Show the plan before execution and let users edit it.From the study
Let users assign individual steps to themselves or the agent.From the study
Allow the plan to change after partial execution.From the study
Ask for help sparingly; it was needed in a minority of tasks.From the study

What they found

Feng et al. (2026)system evaluation

After a formative study with 9 researchers, the authors built Cocoa, a notebook-style agent where plan steps can be assigned to the user or the agent and planning and execution interleave. A lab study (16 people) found it enabled steering without sacrificing ease of use; a week-long field deployment (7 people) showed it used on real research tasks. Builder-run evaluation.

“Cocoa enabled steerability without sacrificing ease-of-use”
Feng et al., 2026 · Abstract

Mozannar et al. (2025)system evaluation

A GPT-4o web agent with co-planning, action approval for irreversible steps, and handover to the user. On 162 GAIA validation tasks, a simulated user with side information raised task completion from 30.3% (autonomous) to 51.9%, while the agent asked for help in only 10% of tasks. A qualitative study with 12 users explored the features; this is a system paper by its builders, not an independent test.

“Magentic-UI implements an action guard system to ensure irreversible or potentially harmful agent actions are reviewed by the human user.”
Mozannar et al., 2025 · Figure 8 caption

Trade-offs and when it doesn't apply

Both are builder-run studies.
Small samples (16 lab, 7 field, 12 users).
Simulated users drove the Magentic-UI gains.

What it looks like

DoAn agent plan with checkboxes: “You do step 3”, “Agent does step 4”.

Don'tA black-box agent that runs to completion with no checkpoints.

How to check it in your product

Track how often users edit plans or take over, and whether those tasks succeed more often.

Open questions

Which steps should default to the user?
How much plan detail helps before it becomes noise?
Do independent evaluations replicate these benefits?

FAQ

Do humans in the loop improve agent results?

In Magentic-UI's evaluation, a simulated user raised task completion from 30.3% to 51.9% on 162 tasks.

Does co-planning make agents harder to use?

Cocoa's lab study (16 people) found it enabled steering without sacrificing ease of use.

Sources

Cocoa: Co-Planning and Co-Execution with AI AgentsK. J. Kevin Feng, Kevin Pu, Matt Latzke, et al. (2026). CHI 2026 (arXiv:2412.10999v4).Evidence type: system evaluation · Quote: Abstract
Magentic-UI: Towards Human-in-the-loop Agentic SystemsHussein Mozannar, Gagan Bansal, Cheng Tan, Adam Fourney, et al. (2025). arXiv:2507.22358 (Microsoft Research, 2025).Evidence type: system evaluation · Quote: Figure 8 captionAlso cited in: When should an AI agent ask for confirmation before acting?