Legion.
View Source
Built-in tool for delegating tasks to sub-agents.
The calling agent must explicitly list allowed sub-agents via tool_config/1:
def tool_config(Legion.Tools.AgentTool), do: [agents: [MyApp.WorkerAgent, MyApp.ResearchAgent]]
def tool_config(_), do: []Only listed agents can be invoked. Attempts to call unlisted agents raise an error.
Usage example from agent code (executed in sandbox)
AgentTool.call(WorkerAgent, "Summarize this data")Listed sub-agents are aliased into the sandbox automatically, so their short names
(the last segment of the module) resolve to the full module atom - no need to spell
out MyApp.Agents.WorkerAgent.
Summary
Functions
Executes a one-off task on a module, or sends a synchronous message to a running agent pid.
Sends a fire-and-forget message to a running agent pid.
Runs multiple sub-agent tasks in parallel and collects results.
Runs sub-agent tasks sequentially, threading each result to the next step.
Starts a long-lived sub-agent process and dispatches task to it asynchronously
via cast/2. Returns {:ok, pid} once the process is started — the task runs
in the background and its result is not returned. Use call/2 if you need the
result, or start_link/2 followed by additional cast/2 calls to queue more
messages on the same process.
Chains a sub-agent call after a previous {:ok, result}. Passes
{:cancel, reason} through unchanged.
Functions
Executes a one-off task on a module, or sends a synchronous message to a running agent pid.
When called with a module, the sub-agent runs to completion and returns
{:ok, result} or {:cancel, reason}. Raises if the agent is not in the allowed list.
When called with a pid, sends a message to the running agent and blocks for the reply.
Sends a fire-and-forget message to a running agent pid.
Runs multiple sub-agent tasks in parallel and collects results.
Returns {:ok, [result1, result2, ...]} when every task succeeds, or the
first {:cancel, reason}. Raises if any agent is not in the allowed list.
Runs sub-agent tasks sequentially, threading each result to the next step.
Each step is {agent, task_or_fn}. If task_or_fn is a 1-arity function,
it receives the previous step's result and must return the task for the
next call. Halts early on the first {:cancel, reason}.
Starts a long-lived sub-agent process and dispatches task to it asynchronously
via cast/2. Returns {:ok, pid} once the process is started — the task runs
in the background and its result is not returned. Use call/2 if you need the
result, or start_link/2 followed by additional cast/2 calls to queue more
messages on the same process.
Raises if the agent is not in the allowed list.
Chains a sub-agent call after a previous {:ok, result}. Passes
{:cancel, reason} through unchanged.