Three levels of responsibility
- The agent — an end-to-end scenario: inputs, stages, tools, a verifiable artifact. Example: the CustDev agent from interviews to an opportunity map.
- The subagent — one function in an existing process: sort emails, compile minutes, find a quote, answer from the database.
- The human — decisions: priorities, sending, legal position, publication. The agent prepares the material, the human makes the decision.
Where the boundary is visible in the product
- The mail assistant sends nothing on its own: the final text and the recipients are confirmed by a human.
- The database consultant has read-only access: writing and modifying data is prohibited at the architecture level.
- The legal RAG navigator forms a package of materials for the lawyer, not a legal opinion.
- The local meeting scribe marks uncertain fragments for verification by the person responsible for the meeting.
Why the boundary is set before launch
Access rights, sources, allowed actions and the confirmation point are fixed before the subagent or agent is connected. This protects both the company and the result: the higher the autonomy and the consequences of a mistake, the stricter the permissions, the journaling and the human confirmation. More in the security section.
How to choose the level
One useful function in a familiar process — a subagent. An end-to-end scenario with an artifact — an agent. Several roles in sequence — a combination like the Golden Trio, where the human checks every handoff.
Questions and answers
Can a subagent grow into an agent?
Yes, if the function grows into an end-to-end scenario: inputs, stages, artifact and acceptance criteria. Then the control requirements change too.
Who is responsible for the agent's mistake?
The result remains material for a human decision: critical conclusions are checked by an employee with the right competence.
Where can I see examples of the boundary?
In the passports: every agent and subagent has its input, output, limits and human control point fixed.
How is a subagent different from an agent?
A subagent performs one simple command on a limited fragment of context; an agent leads a multi-stage process with an artifact.
Can a subagent work without a human?
It returns the result to a human or to the next stage; critical actions are confirmed by an employee.
When do you need an agent rather than a subagent?
When the task requires sequential stages, a journal and a verifiable artifact, not a single step.
Can a process be assembled from several subagents?
Yes: each subagent joins its own step of a familiar process and passes the result on for verification.
Where are the subagent's boundaries fixed?
In the passport: input, output, permissions, perimeter and control points are set before launch.
What happens if a subagent gets extra access?
By the least privilege principle each source is connected separately; access beyond the declared scope is not provided by the interface.
Why is the subagent's price lower?
A subagent is a ready-made function for one step of the process: from $1,500 versus from $6,000 for an agent with the full process. This is the cost of the work: the client pays for the required infrastructure and subscriptions separately; we help with the choice and analysis.