Task Machine vs OpenHands

Give a Goal an agent lead and link Metrics, such as weekly signups or open pipeline value. The lead sees each Metric's latest value and trend, reviews the Goal on its own when a number or the linked work changes and on a schedule, and proposes what is missing for you to approve.

Goals, Projects and Teams each have a lead, a person or an agent. A Goal lead checks that the work can still reach the outcome. A Project lead reviews its Tasks against the Goals that direct it and asks you to settle conflicts between them. A Team lead triages the queue. Every gap becomes a proposal you approve.

Every job is a Task that a person or an agent owns. The Task keeps its plan, discussion, decisions, runs and history together, so you can steer one job in detail without reading a chat log, and people and agents hand work to each other on the same record.

Start a Task or Workflow once, on a schedule in your timezone, or when a service event arrives, such as a paid invoice or a pull request review. A recurring job keeps one stable root Task and gives each occurrence its own status and history, so repeated work does not flood your board.

A Workflow repeats your steps in the same order every time, with branches, human questions and approvals, and verifier checks that judge each result against its requirements and send weak work back with feedback. Every run keeps a step-by-step history you can read afterwards.

Consequential decisions land in your Inbox with their context. You set what each agent may do on its own: every action is either Ask first or Automatic, per agent, and you can change that whenever the work or the stakes change.

Task Machine studies the approvals and corrections you already make. When an agent does well it proposes giving it more freedom, and when it needs closer supervision it proposes tighter review, each with the evidence. A short lucky streak is not enough, and nothing changes until you accept.

Agents keep lessons from your corrections, such as an accepted example or a mistake to avoid, and carry the relevant notes into later work. Memory is kept per agent, project, goal, team and workspace, and you can read, correct and restore any note.

Plan in Chat with your agents, and the plan becomes Tasks or Library documents instead of staying in a chat log. Before an agent starts, you review its plan and the acceptance criteria its work will be judged against.

Send work in over HTTP with a project token. It lands in the project as a Task, or in your Inbox as a proposal first. A team lead agent triages it to the right owner and asks you only when the decision is yours.

Install services from a catalog as MCP servers, APIs or command-line tools, or add your own custom Connector for any of the three. Events Connectors receive webhooks from a service, such as a paid invoice, and start Workflows or wake a waiting Task.

Auto is the default. For each stage of the work it chooses a model from the task, current model capabilities, pricing and your applicable budgets, and you can fix a specific model on an agent when you want to.

Run agents in Managed Cloud with nothing to install, or connect a Local Worker, a computer you control, for local files, tools and accounts. You choose where each agent runs.

Sign in to your ChatGPT (Codex), Claude or GitHub Copilot subscription on a Local Worker, and the agents there use it instead of Task Machine's models.

Usage based. From $29 a month before tax, the whole payment becomes usage that your workspace shares across people and agents, with no base fee and no trial. Larger monthly amounts lower the AI and Cloud rates.

Every Task carries acceptance requirements, and each one gets a Pass, Fail or Uncertain check against the delivered work before a person reviews it. Workflow steps can add a verifier check that sends failed work back with feedback. You review the plan first, so the criteria exist before work starts.

Internal code review checks completed changes against the enabled rules of your repository. Rule discovery reads the repository and proposes rules, each screened against violating and valid examples first, and Task Machine suggests turning off a rule that agents keep dismissing.

When an approved plan has high sensitivity or a Critical review level, the change also gets an adversarial review. Two independent reviewers examine the changed files together for defects such as a broken caller or a duplicate charge on retry. Findings keep their reproduction steps, and a skipped or failed review never counts as a pass.

Before final review, Merge risk rates the exact revision Low, Moderate or High, with notes on alignment, impact, verification, rollback cost and error detection, plus file references. It informs your decision and never merges for you.

Agents deliver against the exact reported revision, address review feedback and update the branch under your repository rules. Task Machine records its own approval of the exact commit before an agent may merge.