An agent lets a person specify a result without planning every step. Delegating that goal creates another question: how far may the AI go to achieve it?
This goes beyond approval for one tool call. A responsibility can span turns, helpers, waiting, and recovery. Individually reasonable actions can still produce cumulative investment or consequences beyond the original intention.
Delegating the steps requires clarity about the goal, action scope, continuing investment, and acceptance. Teloa constrains execution through permissions, allowances, and stopping conditions, leaving human judgment at decisions that change the business.
Make the goal clear and leave room to arrange the steps
Suppose an event budget needs checking. The user asks for costs to be examined, unsupported expenses listed, and a review table delivered. The employee can decide which material to read first, how to organize it, and whether to investigate further without asking the user to direct each step.
More specific delivery requirements make it easier to organize the steps independently. “Take a look at this” gives the employee little basis for deciding when it is done and leaves the person responsible explaining more as work proceeds. A clear result gives judgment a direction without prescribing every action.
A skill supplies the method for checking costs and evidence, the role retains lasting standards, and the task specifies this event and its delivery requirements. With permitted material and available tools, the employee can arrange the checks and collect unsupported or conflicting expenses in one table. The owner judges whether the findings hold and whether to provide more material, without directing which receipt to read next.
Together, these capabilities create room for autonomous work. The goal defines a worthwhile delivery, the method reduces omissions, tools perform reading and organization, and authorization defines where action is permitted. When they hold together, AI can handle a complete stretch of work instead of waiting for each next instruction.
Some steps can proceed independently; others need a decision
Assigning a report involves at least three separate kinds of decision:
- Reading permitted materials and preparing a report stays within the goal and the scope already permitted.
- Sending externally or changing an official business record needs checks along the corresponding execution path. Tool decisions needing the user go through native approval, limited to what was actually approved.
- Accepting the deliverable follows its own delivery check. Accepting a report does not automatically permit sending it externally.
Current role tasks check actual tool calls against the turn’s policy and tool grants, while material access follows the corresponding business scope. For a local business-record change, the owner’s approval refers to a fixed draft and target; adoption checks the version again, and the record can be read afterward. Authorization, execution, and the result each have inspectable evidence.
In the budget review, the employee can continue organizing permitted material. A later step that commits a new budget to a supplier requires checking actual authorization separately. Whether the review table is useful and whether to make an external commitment are two real decisions. Keeping people at those points supports autonomy and responsibility better than confirming every file read, then approving everything together at the end.
Execution also checks current conditions when the owner changes the request or stops the work. Clear authorization reduces repeated confirmation while allowing the mandate to be withdrawn. Waiting and recovery do not erase that change.
Starting another run cannot reset the total investment
A piece of work may pass through a main execution, subtasks, retries, and recovery. If every part treats its allowance as a fresh start, the total investment can grow far beyond the limit the user expected.
Continuing work organized by Teloa currently keeps conditions for turns, tokens, concurrency, elapsed time, retries, and stalled progress. It checks both role-level allowances and the owner’s cumulative allowance. Subwork and recovery remain attributed to the original work; starting another run does not reset the investment.
These allowances constrain execution currently integrated with Teloa’s continuous-work mechanism. Actual monetary charges remain measured by the model and tool services. Both allowance and permission must be valid: remaining capacity still has an action scope, and permission still has an investment limit.
While an outcome is unresolved, the system keeps a reservation pending reconciliation and settles it when actual usage is known. Uncertainty has a cost meaning too; investment awaiting a result remains accounted for.
Bounding investment does not require a person to arrange each step. Goals and skills leave room for execution judgment, while budgets and stopping conditions make that room explainable. Together they allow longer work to be delegated.
Lack of progress also needs a response
Repeated attempts can be useful, or they can circle through the same failure. Ongoing work retains retry and stagnation criteria, so the decision to stop does not depend solely on whether the model is willing to continue.
When those conditions are exhausted, the content already produced, unresolved questions, and corresponding state should remain available. The user can then decide whether to supply materials, change the requirements, or stop. More rounds cannot substitute for diagnosing a problem, and the same gap should not be continually presented as a new step.
This works alongside acceptance of the deliverable. A run can end while the content still needs revision. A budget limits investment without vouching for the quality of the result. People need to see both facts to decide whether to continue.
A stop must reach the actual execution
Stopping should address the original execution and confirm state through its actual response. Closing a detail view only ends inspection. Separating them lets people examine the interface freely and end resource use explicitly. External effects already produced remain subject to actual receipts.
Autonomy and control can strengthen one another. Definite goals reduce step-by-step direction; definite boundaries make delegation easier to trust. The product needs to retain an explainable, revocable working mandate.