Human-in-the-loop coding agents: an alternative to task-to-PR agents
Updated
One kind of product takes a task from the tracker and opens a pull request. CRUX keeps people in the loop at approval checkpoints and works with the coding agents your team already uses.
What a task-to-PR agent is
You assign a task in the tracker to an agent, or an automation rule does it for you. The agent reads the task and the repository, writes the code and opens a pull request. Until then a person usually only watches the session or adds instructions to it.
Task-to-PR agents side by side
| Product | Who starts it | Where it runs | Where a person approves | How the model is paid |
|---|---|---|---|---|
| Claude Agent for Jira | A person assigns the Jira work item to Claude or mentions it in a comment | A managed agent session at Anthropic | A draft GitHub pull request; Claude does not merge | Your own Anthropic API key; the Marketplace app is free |
| Jira Coding Agent | A person assigns the work item to it, or an automation rule runs it | A cloud sandbox at Atlassian | In the session chat, then the person creates the pull request | Rovo credits, at a list price of $0.01 each |
| Linear's cloud coding agent | A person assigns the issue to Linear, or a Loop does when an issue reaches Triage | Linear's cloud, on Claude Code or Codex | A draft pull request in Linear's Reviews tab | AI credits from a prepaid workspace balance, or your own provider key for tokens |
| Cursor Cloud Agents | A person assigns a Linear issue or a Jira work item to Cursor | Isolated virtual machines at Cursor | The pull request the agent opens | Your plan's usage, at the selected model's rate |
| GitHub Copilot cloud agent | A person assigns a GitHub issue to Copilot | A cloud sandbox at GitHub | The pull request Copilot opens, where it asks you for a review | AI credits at $0.01 each |
| Devin | A person assigns a Linear or Jira Cloud task to Devin | A Devin agent session | The pull request Devin opens | A usage allowance in paid plans, extra usage at API pricing |
| Cyrus | A person assigns a Linear issue to it | Your own server, or Cyrus's cloud on the Team plan | Not stated on its site | Your own Claude Code plan |
| CRUX | The person who owns the status starts each step | Each person's own coding agent, or CRUX in the browser | At the agent's understanding, its plan and its result | Each person's own subscription, or the project's model key on CRUX |
Each product name links the page its row comes from, and the last column links the price page. Checked on 13 November 2026.
What this model does well
It suits tasks where the description is the whole task: a typo, a failing test with a clear stack trace, a dependency bump. The work also goes on when nobody on the team is at a keyboard.
Where it stops for a team
For larger tasks the task text is rarely the whole task. Take a made-up task that says "clean up price rounding in checkout". The agent finds a rounding branch that looks redundant and removes it. An earlier task added that branch for one payment provider. Ryan, the reviewer, learns about the misunderstanding in the pull request, when the code already exists.
What a workflow with checkpoints does differently
CRUX starts from what is already in your tracker. Over MCP the agent reads the task, its parent and the decisions already on it, and each status in your workflow has its own instructions. An agent on your machine reads the project wiki from the repository checkout. When it lacks an answer, it asks the person whose role knows it. People approve the agent's understanding, its plan and its result, and a person moves the task to the next status.
The agent runs in each person's own Claude Code, Codex, Cursor or another MCP client, on that person's subscription. You connect it in the dashboard dialog "Connect your coding agent". Nobody acts under a bot account, and text an agent wrote starts with a robot emoji.
When a task-to-PR agent is the better choice
- Most of your tasks are small and fully described.
- You want work to go on when nobody on the team is at a keyboard.
- Your team works in Linear or Jira and wants the agent built into the tracker.
- Nobody on the team has time to answer an agent's questions.
- You want to host the tool yourself, as Cyrus allows.
Limits, price and next step
CRUX is in beta, one founder builds it, and it has no customer stories yet. CRUX is hosted only, with no self-hosted version. It does not support Jira Server, Jira Data Center or GitHub Enterprise Server as a tracker.
One person uses CRUX free. A team of 2 to 10 people pays $49 per active project a month, a team of 11 or more pays $120, and the first month of every project is free. During the beta billing is not switched on, so nothing is charged. The details are on the pricing page.
Start free. Read more in Why CRUX keeps people in the loop. Other comparisons: spec-driven frameworks and Jira MCP and your own skills.
FAQ
What does "human in the loop" mean for coding agents?
A person checks the agent's work at set points. In CRUX the team sets those points in its workflow.
Is CRUX an alternative to Devin?
It is a different model, not a replacement. Devin takes a task and returns a pull request. CRUX runs the coding agents your team already has through your workflow, with people at the checkpoints.
Does a person approve every step in CRUX?
No. People approve at checkpoints, not at every step.
Who pays for the model?
Each person pays through the subscription they already have. An agent started in the browser on CRUX uses the project's own model key.