GenAgent

A behaviour and supervision framework for long-running LLM agent processes, modeled as OTP state machines.

This repository contains six independently versioned Hex packages. Install only the packages an application needs; core has no dependency on a CLI wrapper or HTTP client.

The claude_wrapper and codex_wrapper repositories remain independent.

Each agent is a :gen_statem process wrapping a persistent LLM session. Every interaction is a prompt-response turn, and the implementation decides what happens between turns.

It is a GenServer but every call is a prompt.

GenAgent handles the mechanics of turns. Implementations handle the semantics of turns.

Installation

def deps do
  [
    {:gen_agent, "~> 0.7.0"}, # x-release-please-version
    # Plus at least one backend:
    {:gen_agent_claude, "~> 0.2.0"},
    {:gen_agent_codex, "~> 0.4.0"},
    {:gen_agent_anthropic, "~> 0.3.0"},
    {:gen_agent_openai, "~> 0.3.0"}
  ]
end

Quick start

Define an implementation module by using the GenAgent behaviour:

defmodule MyApp.Coder do
  use GenAgent

  defmodule State do
    defstruct [:path, responses: []]
  end

  @impl true
  def init_agent(opts) do
    path = Keyword.fetch!(opts, :cwd)

    backend_opts = [
      cwd: path,
      system_prompt: "You are a coding assistant."
    ]

    {:ok, backend_opts, %State{path: path}}
  end

  @impl true
  def handle_response(_ref, response, state) do
    {:noreply, %{state | responses: state.responses ++ [response.text]}}
  end
end

Start the agent under the supervision tree and interact with it by name:

{:ok, _pid} = GenAgent.start_agent(MyApp.Coder,
  name: "my-coder",
  backend: GenAgent.Backends.Claude,
  cwd: "/path/to/project"
)

# Synchronous prompt.
{:ok, response} = GenAgent.ask("my-coder", "What does lib/foo.ex do?")
IO.puts(response.text)

# Async prompt.
{:ok, ref} = GenAgent.tell("my-coder", "Add tests for lib/foo.ex")
GenAgent.poll("my-coder", ref)
#=> {:ok, :pending} while the turn runs, then {:ok, :completed, response}

# Or have the outcome sent to you as a message.
{:ok, ref} = GenAgent.tell_with_completion("my-coder", "Run the tests")

receive do
  {:gen_agent, :completion, "my-coder", ^ref, {:ok, response}} -> IO.puts(response.text)
  {:gen_agent, :completion, "my-coder", ^ref, {:error, reason}} -> IO.inspect(reason)
end

# Push an external event into handle_event/2.
GenAgent.notify("my-coder", {:ci_failed, "test_auth"})

GenAgent.stop("my-coder")

One interface for Claude and Codex

Use the same callback module and GenAgent API for both providers. Select the backend when starting each agent:

gen_agent owns the agent process, prompt tasks, callbacks, and queue. gen_agent_claude and gen_agent_codex implement the backend contract, translate provider events, and retain each provider's session identifier. Those backends call claude_wrapper and codex_wrapper, respectively; the wrappers build CLI arguments, parse output, and select a runner for the external process. Add both backend packages to one application when agents need both CLIs. Select a backend per agent, then use the same GenAgent.ask/3, tell/3, poll/3, and interruption APIs for either.

defmodule MyApp.Assistant do
  use GenAgent

  @impl true
  def init_agent(opts) do
    {:ok, [cwd: Keyword.fetch!(opts, :cwd)], %{}}
  end

  @impl true
  def handle_response(_ref, _response, state), do: {:noreply, state}
end

for {name, backend} <- [
      {"claude", GenAgent.Backends.Claude},
      {"codex", GenAgent.Backends.Codex}
    ] do
  {:ok, _pid} = GenAgent.start_agent(MyApp.Assistant,
    name: name, backend: backend, cwd: "/path/to/project"
  )
end

{:ok, response} = GenAgent.ask("claude", "Explain this project")
{:ok, ref} = GenAgent.tell("codex", "Explain this project")
GenAgent.poll("codex", ref) # {:ok, :pending} while queued or running

Either agent supports ask, tell, and poll; the interaction style is independent of the provider. Install both backend packages and configure their CLIs before starting them. Backend-specific options and capabilities can differ: return the appropriate options from init_agent/1 when adding model settings, instructions, or permissions. Each agent keeps its own provider session.

For a runnable OTP application that starts Claude and Codex behind one entry point, see the local GenAgent app example.

State model

An agent is a state machine with two states:

idle --- ask/tell/notify ---> processing
                                  |
                                  v
idle <--- handle_response --- processing (turn done)

• :idle -- waiting for work. On enter, drains the mailbox (queued prompts) in FIFO order.
• :processing -- a prompt is in flight. One at a time. New prompts queue.
• Self-chaining -- handle_response/3 can return {:prompt, text, state} to immediately dispatch another turn without going through the mailbox. Useful for multi-step work that the agent drives itself.
• Halting -- handle_response/3, handle_error/3, handle_event/2, or pre_turn/2 can return {:halt, state} to go idle but freeze the mailbox. A halted agent ignores queued prompts until GenAgent.resume/1 is called.
• Watchdog -- a :state_timeout kills any turn that runs longer than the :watchdog_ms deadline (positive integer milliseconds or :infinity, default 10 minutes). :max_tell_results (non-negative integer, default 100) bounds how many tell/2 results poll/2 retains; zero retains none.

Lifecycle hooks

In addition to the core callbacks, v0.2 adds four optional lifecycle hooks for fine-grained control over what happens around each turn and around the agent's full run:

All four are optional with default no-op implementations. The guiding principle is telemetry first, callbacks for state mutation -- observational use cases (log tokens, emit metrics) should use the existing telemetry events; callbacks exist specifically for hooks that need to mutate agent_state or block the next transition.

See the Workspace pattern guide for a complete example exercising all four hooks in sequence around a git workspace.

Backends

GenAgent ships with a GenAgent.Backend behaviour and no built-in backend. Pick one of the sibling packages or write your own:

A backend owns its session lifecycle, translates the LLM-specific event stream into the normalized GenAgent.Event values the state machine consumes, and carries any state it needs (session id, message history) in an opaque session term.

The contract has three required callbacks (start_session/1, prompt/2, terminate_session/1). prompt/3 and checkpoint_session/2 let CLI backends acknowledge a session ID during an active turn, so a later failure or interruption does not discard it. update_session/2 and resume_session/2 remain optional. See GenAgent.Backend for details.

See the Backends guide for backend implementation rules and a capability and option comparison.

Public API

Names resolve through a Registry, so callers address agents by name (any term). Agents use restart: :temporary: a crashed or stopped agent must be started explicitly, and its previous state is not restored.

runtime_snapshot/2 reports the coordinator's current phase, halted flag, queued prompt and buffered notification counts, pending self-chain, and active request ref, origin and elapsed/watchdog time. It contains no prompt, callback state, backend session or event payload. It is a point-in-time observation, not durable state or permission to dispatch. The older status/2 API remains available; its agent_state is the server's latest retained state, not a live read of an in-flight task.

Request completion messages

Use tell_with_completion/4 when a caller needs the exact request ref and an asynchronous terminal outcome:

{:ok, ref} = GenAgent.tell_with_completion("my-coder", "Run the tests")

receive do
  {:gen_agent, :completion, "my-coder", ^ref, {:ok, response}} ->
    IO.puts(response.text)

  {:gen_agent, :completion, "my-coder", ^ref, {:error, reason}} ->
    IO.inspect(reason)
end

The optional third argument selects a recipient pid; it defaults to the caller. The agent registers that recipient when it accepts the request, so a fast response or pre_turn/2 skip can send the completion before the call returns. Match on the ref, which works with interrupt_request/3 while active or cancel_request/3 while queued. Pending-queue overload returns an error without a ref or completion message. Accepted queued requests deliver after their turn, or deliver {:error, :cancelled} when cancelled before dispatch. Backend errors, gate skip/halt/invalid, interruption, and watchdog expiry deliver error outcomes. For started turns, the message is sent after turn decision and post_turn/3 callbacks. It is sent at most once for each accepted request.

Delivery does not depend on the bounded poll-result cache. It is an in-memory BEAM send, not durable delivery: a dead recipient loses its message, and agent death before a terminal outcome leaves the request uncertain. Monitor the agent if its death matters; a replacement under the same name uses new refs. A crashing decision callback stops the agent before completion. Neither a completion nor an agent monitor proves that an external provider process has settled.

Pending input bounds

Each agent admits at most 1,000 pending prompts and 1,000 deferred notifications by default, with a separate 1,048,576-byte payload cap for each queue. Configure :max_pending_prompts, :max_pending_prompt_bytes, :max_pending_notifications, and :max_pending_notification_bytes in start_agent/2 or child_spec/2. Values must be non-negative integers; zero disables the corresponding pending queue. Byte use is the sum of :erlang.external_size/1 for queued prompt strings or event terms. The active prompt and events handled immediately while idle are outside these pending caps.

ask/3 and tell/3 return {:error, {:overloaded, info}} before acceptance when a pending prompt does not fit. A rejected tell/3 returns no pollable ref. info identifies the queue, count or byte limit, current count and bytes, incoming bytes, and configured maxima. notify_ack/3 returns :ok after the event is handled or retained, or the same typed overload error if admission fails. It acknowledges in-memory processing, not durable delivery. Existing notify/2 is a best-effort cast: its immediate :ok does not mean the event fit. Use the rejection telemetry event to observe cast overloads.

Deferred event callbacks may generate prompts when a turn finishes. Those prompts use the bounded pending-prompt queue; on overload, handle_error/3 receives the reason. Self-chain prompts from handle_response/3 or handle_error/3 use one reserved slot outside the prompt count cap, but must fit the prompt byte cap. A halted agent retains admitted work until resume/1; interruption or completion releases queue capacity as work is drained. These limits bound accepted internal storage, not arbitrary messages already waiting in the BEAM process mailbox before admission is processed. Stream-output capture and external provider cancellation have their own contracts.

Event capture bounds

Each turn retains at most 1,000 normalized events and 1,048,576 bytes of event terms by default. Override these positive-integer limits with :max_events_per_turn and :max_event_bytes_per_turn in start_agent/2 or child_spec/2. Byte usage is the sum of :erlang.external_size/1 for retained events, not total process memory.

The default event_retention: :compact delivers every normalized event to handle_stream_event/2 and continues to the terminal event. If the log limit is reached, Response.events holds the unchanged prefix and Response.event_coverage reports the observed, retained, and omitted counts plus the first omission reason. Response.terminal keeps the full terminal event separately; Response.text, usage, and session ID are built from the full stream. Check coverage before using Response.events as a complete record. Full response text, terminal payloads, callback state, and provider decoder buffers can still grow beyond the event-log budget.

Set event_retention: :lossless to require the previous complete-log contract. An event that exceeds either limit then ends the turn with {:error, {:event_capture_overflow, diagnostics}}; the rejected event is not delivered to handle_stream_event/2. The stream enumerable is halted, but settlement of provider subprocesses or remote work remains the backend's responsibility.

Supervision

The package starts a fixed supervision tree on application boot:

GenAgent.Supervisor
  GenAgent.Registry          (Registry, keys: :unique)
  GenAgent.TaskSupervisor    (Task.Supervisor)
  GenAgent.AgentSupervisor   (DynamicSupervisor)
    <your agents under here>

To own agent and prompt-task lifetimes in your application, start a Task.Supervisor before a DynamicSupervisor in your supervision tree:

children = [
  {Task.Supervisor, name: MyApp.AgentTasks},
  {DynamicSupervisor, name: MyApp.Agents, strategy: :one_for_one}
]

{:ok, _owner} = Supervisor.start_link(children, strategy: :rest_for_one)

spec =
  GenAgent.child_spec(MyAgent,
    name: "worker-1",
    backend: MyBackend,
    task_supervisor: MyApp.AgentTasks
  )

{:ok, _agent} = DynamicSupervisor.start_child(MyApp.Agents, spec)
{:ok, response} = GenAgent.ask("worker-1", "Hello")
:ok = GenAgent.stop("worker-1", MyApp.Agents)

child_spec/2 requires an explicit, running task supervisor; it never silently uses the global one. Both globally and caller-owned agents use GenAgent.Registry, so names must be unique across them and normal name-based calls work for either. stop/1 targets only the global agent supervisor; pass the caller's DynamicSupervisor to stop/2. The agent child is temporary and is never automatically replayed after a crash. With :rest_for_one, failure of the task supervisor also stops the agent supervisor. On ordinary shutdown, the agent supervisor stops first so its agents can cancel in-flight tasks while their task supervisor is still running. An application must reconstruct state and consumed work from its own durable records; supervision alone does not provide recovery. The Chaos Lab example is a keyless, executable check of the supervision contract.

Each prompt turn runs as a Task under its selected Task.Supervisor. A crashed task delivers :DOWN to the owning agent, which turns it into an {:error, {:task_crashed, reason}} response for the caller -- it does not take down the agent process. If the selected task supervisor is unavailable before a task can start, the request fails with :task_supervisor_unavailable, reaches handle_error/3, and leaves the agent idle. A callback's immediate retry is discarded in this case; a later request can run after the supervisor returns.

The prompt task belongs to its agent: it is stopped when the agent exits, including abrupt exits that bypass termination callbacks. Interruption and the watchdog also stop the active prompt task. Stopping a BEAM task does not establish that a provider's subprocess or remote request has stopped; external cancellation and resource cleanup belong to the backend and its transport.

Patterns

Eleven common topologies are documented as ex_doc guides shipped with the package. Each guide is a complete worked example you can read, copy, and adapt -- they are not installed as public API modules:

• Switchboard -- human-managed named agent fleet with non-blocking send/poll/inbox, the base for manager-driven UIs
• Research -- one agent self-chaining through phases
• Debate -- two agents pushing each other via cross-notify
• Pipeline -- linear stage chain, one-way notify
• Supervisor -- coordinator + dynamic workers (fan-out/in)
• Pool -- reusable worker pool with round-robin dispatch
• Watcher -- reactive event-driven agent, idle until triggered
• Heartbeat -- periodic agent driven by timer events
• Checkpointer -- human-in-the-loop review workflow
• Retry -- handle_error self-chain for transient failures
• Workspace -- all four lifecycle hooks around a git workspace

Start with the patterns overview for a "choose your pattern" decision tree.

Telemetry

GenAgent emits telemetry events for observability:

[:gen_agent, :turn, :start]      # content-free, %{agent, ref, origin}
[:gen_agent, :turn, :stop]       # content-free, duration_ms
[:gen_agent, :turn, :error]      # content-free, duration_ms, reason_kind
[:gen_agent, :turn, :rejected]   # declined before dispatch; no matching start
[:gen_agent, :prompt, :start]    # %{agent, ref}
[:gen_agent, :prompt, :stop]     # %{agent, ref, duration}
[:gen_agent, :prompt, :error]    # %{agent, ref, reason}
[:gen_agent, :event, :received]  # %{agent, event}
[:gen_agent, :state, :changed]   # %{agent, from, to}
[:gen_agent, :mailbox, :queued]  # %{agent, depth}
[:gen_agent, :input, :rejected]  # %{agent, reason: {:overloaded, info}}
[:gen_agent, :halted]            # %{agent}

Enough to build a communication graph, track latency, alert on stuck agents. Attach handlers with :telemetry.attach/4. Use the :turn events for ordinary metrics: older prompt/event events can carry prompts, raw errors, or agent state. GenAgent.Telemetry documents measurements, units, ordering, and data handling.

What GenAgent does not do

• Prescribe agent behavior. No retry logic, no STATUS line conventions, no summary format. That is all implementation concern.
• Prescribe inter-agent communication. Agents can notify/2 each other by name, but the message format is up to you.
• Manage persistence. If you want to persist agent state across restarts, do it in terminate_agent/2 and init_agent/1.
• Manage pools. One agent = one session = one process. If you want a pool, start multiple and route to them.
• Track costs or budgets. Usage data is in GenAgent.Response.usage. Do what you want with it.

Testing

mix test
mix format --check-formatted
mix credo --strict
mix dialyzer

These commands check the root package. Run scripts/quality.sh to validate all six packages; see the contributor guide for setup and release conventions.

The test suite uses an in-process GenAgent.Backends.Mock (in test/support/) that lets you script backend responses without any external process. See test/gen_agent/server_test.exs for examples.

License

MIT. See LICENSE.