Create a Custom Human Task

Custom Task is useful when software knows what information or action it needs but the final step requires a person at a physical location. On HumanTask API, the request is framed as a bounded job: create a one-off human request when the physical action does not fit an existing standardized capability. The requester defines the target, timing and proof before anyone accepts the work.

When to Use Custom Task

The best custom task jobs are concrete enough that two independent workers would understand the same objective. Examples include observe a specific condition; perform a simple approved action; collect a niche piece of local information; or combine several small steps at one location. These are not broad consulting assignments. They are observable field actions that can be accepted, completed and checked.

Define a Custom Task Request

A requester should provide clear objective, location, and step-by-step instructions. It should also state deadline, budget, and evidence and stop conditions. The instruction should separate facts the worker can observe from decisions the AI or business will make later. That distinction prevents a simple field task from quietly turning into specialist advice.

Required inputs

• Clear objective
• Location
• Step-by-step instructions
• Deadline
• Budget
• Evidence and stop conditions

Evidence That Makes the Result Useful

For custom task, a useful completion package may contain task-specific proof, structured answers, media where relevant, timestamp, and worker notes. The evidence schema should be selected before dispatch. A task that merely says “send proof” is weaker than one that specifies which files, fields and observations are mandatory. The receiving agent can then test completeness without interpreting a chat message.

Suggested result fields

• Task-specific proof
• Structured answers
• Media where relevant
• Timestamp
• Worker notes

Failure Modes to Plan For

Real locations create exceptions that software APIs rarely face. In this capability, common examples are: scope is too broad, the request requires licensed expertise, instructions are unsafe or impossible, and success cannot be verified objectively. Those outcomes should be returned as explicit exception states rather than hidden inside a free-text note. An AI agent can then retry with new instructions, choose another location, widen the deadline or escalate to a specialist.

Exception examples

• Scope is too broad
• The request requires licensed expertise
• Instructions are unsafe or impossible
• Success cannot be verified objectively

Example Workflow for an AI Agent

Consider this workflow: A team can test an unusual workflow manually as a custom task, then promote it into a standardized capability if demand repeats. The agent first decides that a physical check is necessary, then creates a custom_task task with the address, deadline and evidence fields. The worker accepts the job, completes only the permitted actions and submits the requested proof. After the result arrives, the software can validate required fields, store the media references and continue its original plan.

Who Uses Custom Task

Custom Task can support AI agents, developers, researchers, remote operations and businesses testing new workflows. These users have different business goals, but they share the same bottleneck: the missing fact or action exists offline. A reusable task definition lets them solve that bottleneck without maintaining a field team in every city.

API Shape for Custom Task

The machine-facing representation should be narrow. A custom_task request can carry a normalized location, human-readable instructions, a deadline, budget, required evidence and a client reference. The response should return a task identifier and lifecycle state. Follow-up operations should expose status and evidence without forcing the caller to scrape a dashboard. MCP can present the same operation as an agent tool; REST and OpenAPI can serve conventional application integrations.

Example task object

{
  "capability": "custom_task",
  "location": {"address": "TARGET_ADDRESS"},
  "deadline": "ISO_8601",
  "instructions": "TASK-SPECIFIC_INSTRUCTIONS",
  "evidence_required": ["TASK_SPECIFIC_FIELDS"]
}

Launch a Custom Task Task

Start with one custom task request that has an unambiguous outcome. Define the location, deadline and proof first; then create the task through the web flow or the available developer interface. If the workflow repeats, promote the same evidence schema into a reusable integration.

What makes custom task different from a generic gig

The value is not simply that a person is available. The value is that the request is standardized enough for software to understand the expected result. For custom task, the schema should reflect the actual decision being supported: the agent needs evidence about create a one-off human request when the physical action does not fit an existing standardized capability. That makes the task easier to price, route, compare and audit than an open-ended message to a freelancer.