Agent-to-Agent Delegation
What agent-to-agent delegation is
Most organizations end up with more than one agent, each good at something different — one that knows the sales pipeline, one that knows billing, one that knows a technical product inside out. Before this, an agent that met a question outside its own area could only answer from general knowledge or tell the person to ask someone else.
With delegation, the first agent (the calling agent) can put the question to a second agent (the consulted agent) in the middle of the conversation, wait for the answer, and fold it into its own reply. The person chatting stays in one conversation the whole time.
The consulted agent answers using its own instructions, its own tools and its own data access — that is the point. You do not have to copy one agent's knowledge into another to make it available.
What the consulted agent can and cannot do
A consultation is deliberately limited. It is a question-and-answer, not a handover of control.
| Behaviour | What it means for you |
|---|---|
| Runs as the person using the assistant | The consulted agent sees exactly what that person could already see — no more. Delegation never widens anyone's data access. |
| Cannot change anything on its own | If the consulted agent proposes an action — creating a record, sending an email — that proposal comes back into the original conversation for the person to confirm. It is never carried out silently. |
| Kept as its own conversation | Each consultation is recorded separately, so you can see afterwards what was asked and what came back. Useful for auditing and for checking whether the specialist agent is answering well. |
| Costs an extra turn | A consultation is a second agent doing real work, so it consumes credits in addition to the original conversation. Depth (below) is how you keep that bounded. |
| Only happens when instructed | Switching the feature on does not make agents start consulting each other. The calling agent's instructions have to tell it when a question belongs to a specialist. |
Turning it on
Two settings have to be switched on, and they are on the agent that will answer, not the one that asks. Both start off, so an upgrade never changes how your existing agents behave.
- Go to App Launcher → REDA AI → Agent Definitions and open the agent you want to make available for consultation — the specialist.
- In the Agent Collaboration section, tick Agent Invocable. This is what makes the agent visible to other agents at all.
- On the same record, make sure Allow Automated Requests is also ticked. A consultation arrives as an automated request, so both are required. If only one is on, the calling agent gets a message naming the setting that is still missing — it will not fail silently.
- Save.
- Now open the agent that should be able to ask — the generalist — and set Max Agent Call Depth (see below). Leave it at
1unless you have a reason to change it. - Finally, tell the calling agent when to consult, in its instructions. For example: “For detailed technical questions about installed equipment, consult the Technical Specialist agent and summarise its answer.” Without this the capability sits unused.
How deep consultations can go
Max Agent Call Depth is set on the agent that starts the conversation, and it limits how many levels of consultation one question may trigger.
| Value | What happens | When to use it |
|---|---|---|
0 |
This agent never consults anyone. | An agent you want kept strictly self-contained. |
1 (default) |
It may consult one agent. That agent answers directly and cannot consult a third. | Almost always. One hop covers the “ask the specialist” case with predictable cost and speed. |
2 or more |
Consultations may chain — the agent it consults may itself consult another. | Rarely. Each level adds response time and credits, so raise it only for a case you have actually seen need it. |
Left blank, the field behaves as 1. An agent can never consult itself, and a chain that would loop back to an agent already in it is stopped — so a circular set of instructions cannot run away.
What the person in the conversation sees
The consultation is not shown as a separate step in the transcript. The person sees the calling agent's reply, which contains the answer, and an indicator that a tool was used.
If you want people to know which agent was consulted — often worth it, because it explains where an answer's authority comes from — instruct the calling agent to say so. For example: “When you consult another agent, begin your reply by naming it.” The agent will then open with something like “I checked with the Technical Specialist —” before giving the answer.
When to use it, and when not to
A good fit: you have genuine specialist knowledge that lives with one agent — a product, a regulation, a process — and other agents keep running into questions about it. Delegation lets that knowledge stay in one place, maintained once, and still be reachable from everywhere.
Not a good fit: two agents that need the same background information. Duplicating a short set of instructions is simpler and cheaper than a consultation on every conversation. Delegation earns its cost when the specialist is genuinely deep, not when it is merely a paragraph the other agent could have carried itself.
Why a consultation did not happen
When a consultation cannot go ahead, the calling agent is told why and carries on with what it already knows, rather than failing the conversation. These are the reasons you will see.
| What the agent reports | What to do |
|---|---|
| The agent is not available to be consulted | Tick Agent Invocable on that agent. |
| The agent is not set up to answer requests from other agents | Tick Allow Automated Requests on that agent. Both settings are needed. |
| No agent was found with that id | The calling agent's instructions name an agent that does not exist. Check the spelling against the Agent Definitions list. |
| The agent is not active | Activate the specialist agent, or point the instructions at one that is active. |
| The consultation limit has been reached | Expected when a chain hits Max Agent Call Depth. Raise it only if the chain is genuinely needed. |
| There is not enough processing capacity left in this request to consult another agent | Happens on very long or heavy conversations — the agent answers with what it already has. No configuration change is needed; if it recurs, shorten the conversation or ask the specialist directly. |
| An agent cannot consult itself | The instructions point the agent at its own name. Point them at the specialist instead. |