Hayata

Hayata is an Erlang/OTP framework for building AI agents around a graph-based execution model. It uses OTP processes, supervision, monitoring, and message passing to coordinate agent execution, with explicit ownership of process state and durable recovery checkpoints.

Hayata is designed to be usable from Erlang and Elixir applications. Its graph-oriented model is inspired by agent orchestration frameworks such as LangGraph, while exposing an Erlang/OTP-native runtime rather than a feature-for-feature clone of another API.

Highlights

• Graph-based agent execution with a configurable decision policy.
• Sequential and parallel node execution.
• OTP supervision and process monitoring for agents, nodes, edges, reducers, and edge watchers.
• Asynchronous invocation APIs and result notifications.
• Invocation checkpoints and task-result persistence in Mnesia.
• Recovery of interrupted invocations, including parallel transitions.
• Conservative handling of uncertain tasks: a task marked as running without a durable result is not silently re-executed.
• Telemetry lifecycle events for agents, nodes, edges, and reducers.
• Provider-agnostic RAG retrieval with normalized document results, usable directly from nodes or through LLM tool calling.

Architecture

• Agent --- coordinates invocations, applies the decision policy, tracks invocation state, and initiates recovery.
• Node --- a supervised gen_server that owns local execution state, coordinates outgoing edges, tracks executions, and monitors edges and reducers.
• Edge --- an anonymous gen_server that runs an edge task and reports its terminal result.
• Edge watcher --- monitors the node and its edge. If the node terminates unexpectedly, the watcher terminates the edge so it does not remain orphaned.
• Reducer --- a temporary process that collects edge results for an execution and produces a consolidated outcome.
• Task store --- persists invocation checkpoints and task outcomes in Mnesia.
• Node supervisor --- supervises node processes and restarts them after abnormal termination.

Requirements

• Erlang/OTP 29.1.1 (target runtime)
• Rebar3

Build and test

rebar3 compile
rebar3 eunit

Quick start

The following example mirrors a basic graph-agent pattern: define a node, define control flow, start the agent's supervision tree, and invoke it. This example uses a mock node; it does not call an LLM.

%% hello_hayata.erl
-module(hello_hayata).

-export([run/0]).

run() ->
    AgentId = hello_agent,

    MockNode = fun(State) ->
        Message = maps:get(input, State, <<"Hello">>),
        #{
            messages => [
                #{role => ai, content => <<"Hello, world!">>},
                #{role => user, content => Message}
            ]
        }
    end,

    Nodes = #{mock_node => MockNode},

    DecisionPolicy = fun(Context) ->
        case maps:get(step, Context) of
            0 ->
                {next, mock_node, <<"Hello">>};
            1 ->
                {finish, maps:get(state, Context)}
        end
    end,

    Config = #{
        id => AgentId,
        state => #{},
        nodes => Nodes,
        decision_policy => DecisionPolicy,
        max_steps => 3,
        execution_timeout => 5000
    },

    case hayata_sup:start_agent(Config) of
        {ok, AgentSupPid} ->
            try
                run_agent(AgentSupPid)
            after
                hayata_sup:stop_agent(AgentSupPid)
            end;
        {error, Reason} ->
            {error, {agent_start_failed, Reason}}
    end.

run_agent(AgentSupPid) ->
    case hayata_agent_sup:agent(AgentSupPid) of
        {ok, AgentPid} ->
            case hayata_agent:run(AgentPid, #{input => <<"Hello">>}, #{subscriber => self()}) of
                {ok, InvocationId} ->
                    await_result(InvocationId, 100);
                {error, Reason} ->
                    {error, Reason}
            end;
        {error, Reason} ->
            {error, Reason}
    end.

await_result(_InvocationId, 0) ->
    {error, notification_timeout};
await_result(InvocationId, Attempts) ->
    receive
        {hayata_agent_result, InvocationId, {ok, Result}} ->
            {ok, Result};
        {hayata_agent_result, InvocationId, {error, Reason}} ->
            {error, Reason}
    after 100 ->
        await_result(InvocationId, Attempts - 1)
    end.

Save the module as hello_hayata.erl in a location included in your Erlang code path, compile it, and call:

hello_hayata:run().

Agent Execution Contract

Hayata uses two different function contracts during agent execution:

• Node functions receive the node execution state and return a map of state updates.
• Decision policies receive the agent execution context and decide which node executes next, whether multiple nodes should execute in parallel, or whether the invocation should finish.

Both functions receive a single map argument, but their available fields and responsibilities are different.

Node Function

Every node is defined as a one-argument function:

Node = fun(State) ->
    %% Read the current state.
    Input = maps:get(input, State),

    %% Return state updates.
    #{
        answer => Input,
        status => completed
    }
end.

The State argument contains the agent's current shared state, with the current node input added under the input key.

Its structure is:

#{
    %% All fields from the agent's current shared state.
    input => NodeInput,

    %% Other application-defined state fields.
    messages => [],
    customer => undefined
}

The following fields are available:

Returning node updates

A node must return a map containing the state updates it wants to publish:

Node = fun(State) ->
    Input = maps:get(input, State),

    #{
        response => Input,
        processed => true
    }
end.

Hayata merges the returned map into the current shared agent state. Existing keys with the same name are replaced by the returned values; keys not included in the update remain unchanged.

NewState = maps:merge(CurrentState, NodeUpdates).

Nodes do not directly mutate the shared agent state. They return updates, and Hayata applies them after successful execution.

Node input and state

The input passed to a node is determined by the decision policy:

{next, NodeId, NodeInput}

Hayata adds this value to the current shared state under the input key before invoking the node.

For parallel execution, every selected node receives the same input and the same current shared state. Their results are collected independently and are not merged into shared state automatically.

Partial results

When invocation streaming is enabled, a node can publish partial results using hayata_streamer:

Node = fun(State) ->
    Streamer = maps:get(hayata_streamer, State),

    ok = Streamer(#{chunk => <<"First part">>}),
    ok = Streamer(#{chunk => <<"Second part">>}),

    #{completed => true}
end.

Partial results are emitted as node_partial_result events. They are notifications and do not modify shared agent state.

Decision Policy

The decision policy is also a one-argument function:

DecisionPolicy = fun(Context) ->
    case maps:get(step, Context) of
        0 ->
            {next, first_node, <<"Hello">>};
        1 ->
            {finish, maps:get(state, Context)}
    end
end.

The Context map provides information about the current agent invocation and its execution progress.

Its structure is:

#{
    agent_id => AgentId,
    state => CurrentSharedState,
    input => OriginalInvocationInput,
    step => CurrentStep,
    max_steps => MaxSteps,
    parallel_results => ParallelResults
}

The following fields are available:

Policy decisions

The decision policy must return one of the following values:

Execute one node

{next, NodeId, Input}

Schedules a single node and passes Input to its function.

Execute multiple nodes in parallel

{next, [NodeA, NodeB], Input}

Schedules multiple distinct nodes concurrently. All nodes receive the same input and shared state.

When every node completes successfully, their returned update maps become available through parallel_results, preserving the order of the requested node identifiers.

Finish the invocation

{finish, Result}

Completes the invocation and publishes Result as its final outcome.

Return an error

{error, Reason}

Fails the current invocation with the provided reason.

Any other return value is considered an invalid policy decision.

Parallel execution example

DecisionPolicy = fun(Context) ->
    case maps:get(step, Context) of
        0 ->
            {next, [customer_node, orders_node], maps:get(input, Context)};

        2 ->
            [Customer, Orders] = maps:get(parallel_results, Context),

            {next, aggregate_node, #{
                customer => Customer,
                orders => Orders
            }};

        3 ->
            {finish, maps:get(state, Context)}
    end
end.

The parallel_results list follows the order in which the nodes were requested, regardless of their completion order.

Important distinctions

This separation keeps node execution focused on processing data and returning state updates, while the decision policy remains responsible for orchestration and control flow.

Conversational memory

Agents can optionally preserve conversation history between successive invocations. Enable this behavior with conversation_memory => true in the agent configuration:

Config = #{
    id => support_agent,
    state => #{messages => []},
    nodes => Nodes,
    decision_policy => DecisionPolicy,
    conversation_memory => true
}.

When conversational memory is enabled, Hayata carries the messages entry from the agent's current state into the next invocation. This lets a node build each LLM request from the previous conversation instead of starting with an empty message history every time.

For example, if a node appends each new user message to the messages list, successive invocations can build a history like [FirstMessage, SecondMessage]. The application is responsible for appending messages in the format expected by its LLM provider and for deciding which messages belong in the conversation.

The option defaults to false. With the default behavior, every invocation starts from the configured initial state, so message history from a previous invocation is not retained. When the option is enabled, only the messages entry is carried over; other state fields are reset to their configured initial values at the beginning of each invocation.

This feature preserves history in the lifetime of the running agent. It is not a vector or semantic memory system, and it does not by itself provide a durable conversation store or guarantee that conversation history survives an agent or VM restart.

Invocation API

Hayata supports asynchronous agent invocations with optional result notifications and event streaming.

Starting an invocation

Start an invocation without a subscriber:

{ok, InvocationId} = hayata_agent:run(AgentPid, Input).

The invocation runs asynchronously. Use hayata_agent:get_result/1 to retrieve the final outcome when it is available.

To receive a terminal result notification, provide a subscriber PID:

{ok, InvocationId} = hayata_agent:run(
    AgentPid,
    Input,
    #{subscriber => self()}
).

The subscriber receives one of these messages:

{hayata_agent_result, InvocationId, {ok, Result}}
{hayata_agent_result, InvocationId, {error, Reason}}

InvocationId identifies the invocation. If the invocation is rejected before it starts, run/3 returns an error instead of sending a terminal result notification for that request.

Streaming events

Enable streaming by setting stream => true and providing a subscriber:

{ok, InvocationId} = hayata_agent:run(
    AgentPid,
    Input,
    #{
        subscriber => self(),
        stream => true
    }
).

The subscriber receives streaming events in this format:

{hayata_agent_event, InvocationId, Event}

Each Event is a map containing:

• type — the event type.
• sequence — a monotonically increasing sequence number within the invocation.
• data — event-specific information.

The event types are:

Sequence numbers start at 1 for each invocation. In parallel transitions, lifecycle and partial-result events from different nodes may interleave. Use node_id and execution_id to associate events with their corresponding executions.

Streaming events do not replace the terminal result notification. When streaming is enabled, the subscriber receives both hayata_agent_event messages and the traditional hayata_agent_result notification.

Emitting partial results from a node

Node functions receive a context map as their single argument. When streaming is enabled, Hayata injects a hayata_streamer function into that context.

The streamer accepts a map and emits a node_partial_result event:

Node = fun(Context) ->
    Streamer = maps:get(hayata_streamer, Context),

    ok = Streamer(#{chunk => first}),
    ok = Streamer(#{chunk => second}),

    #{answer => 42}
end.

A partial-result event has this shape:

#{
    type => node_partial_result,
    sequence => 3,
    data => #{
        node_id => answer_node,
        execution_id => ExecutionId,
        edge_id => node_result,
        result => #{chunk => first}
    }
}

The example sequence number is illustrative. Actual sequence numbers depend on the other events emitted during the invocation.

The streamer accepts maps only. Passing a non-map returns {error, invalid_partial_result}. A partial result from an execution that is no longer active is rejected rather than emitted as a valid event.

Partial results are notifications; they do not independently modify the agent's shared state or replace the node's final result. The node must still return its final result map as usual.

The streamer synchronously submits the partial result to the agent. This allows the agent to publish the partial event before processing the node's terminal completion notification.

Streaming disabled

Streaming is disabled by default. Existing invocations can continue to use run/2 or subscribe to terminal results without enabling streaming.

When streaming is disabled, hayata_streamer is not injected into the node context and no hayata_agent_event messages are emitted.

Run options

hayata_agent:run/3 accepts a map with these supported keys:

Streaming requires a subscriber. Unknown options, an invalid subscriber, or an invalid stream value return an error.

Parallel node execution

A decision policy can schedule multiple nodes in parallel by returning {next, NodeIds, Input}, where NodeIds is a non-empty list of distinct node identifiers.

All branch nodes receive the same input and current shared agent state. Their returned maps are collected independently and do not directly modify the shared state.

When every branch completes successfully, the policy receives the results through the parallel_results field in its context. Results are a list of maps in the same order as the requested node identifiers, regardless of completion order.

DecisionPolicy = fun(Context) ->
    case maps:get(step, Context) of
        0 ->
            {next, [customer, orders], undefined};
        2 ->
            [Customer, Orders] = maps:get(parallel_results, Context),
            {next, aggregate, #{
                customer => Customer,
                orders => Orders
            }};
        3 ->
            {finish, maps:get(state, Context)}
    end
end.

The parallel_results field is undefined before a parallel transition completes. After all branches succeed, it contains their ordered result maps and remains available to the policy that chooses the next transition. When the policy schedules another node or parallel transition, the previous results are cleared. A new invocation starts with parallel_results set to undefined.

If a branch fails or the invocation times out, the transition fails and pending branches are cancelled; incomplete parallel results are not published as a successful transition result.

Durable checkpoints and recovery

Hayata uses Mnesia to persist invocation checkpoints and task outcomes. The task store uses disk-backed Mnesia tables and synchronous transactions; completed task results and invocation checkpoints are flushed to the Mnesia log after writes.

A checkpoint captures the information required to identify and resume an invocation, including the invocation ID, current state and step, active execution, parallel results, and terminal result/error fields. When an agent starts and finds an unfinished invocation for its ID, it restores that invocation and reports recovery_required rather than accepting a new invocation on top of it.

Call hayata_agent:recover(AgentPid) to initiate recovery of the restored invocation.

Recovery classifies each task using its durable record:

• Completed: reuse the persisted task result instead of executing that task again.
• Not started: schedule the task for execution during recovery.
• Uncertain: if a task record indicates that execution started but no completed result exists, recovery returns {error, task_execution_uncertain} rather than silently repeating the operation.

For parallel execution, Hayata can combine completed branch results with branches that have not started, execute only the missing branches, and persist the final invocation result. Results retain the configured node order.

Recovery guarantees and limitations

Recovery is designed to avoid silently repeating a task whose completion is uncertain. It does not provide exactly-once semantics for arbitrary external side effects. If a node may have called an external service before the VM stopped but its result was not durably saved, Hayata cannot know whether that side effect occurred. Applications should use idempotency keys, external deduplication, or other application-level safeguards where retry safety matters.

Durable recovery depends on preserving the configured Mnesia data directory across VM restarts. If that directory is removed or replaced, the persisted checkpoints are unavailable.

Reproducing VM restart scenarios

The repository includes VM_RECOVERY_TESTS.md, a step-by-step manual verification guide with complete local commands for:

1. A completed task result surviving an abrupt VM shutdown.
2. A task left running without a result being classified as uncertain after restart.
3. A running invocation checkpoint being restored by a new agent process.
4. A parallel invocation reusing a completed branch and executing a not-started branch after an abrupt VM shutdown.

These manual integration checks complement the EUnit suite; run both against the exact commit intended for release.

Execution lifecycle

1. The node validates an execution request, edge tasks, and timeout.
2. A reducer is started and activated when result consolidation is required.
3. The node starts and monitors edge processes.
4. Each edge runs its task and sends its result to the reducer and node.
5. The reducer consolidates the edge results and reports the final result.
6. The node updates execution status, notifies any configured recipient, and tracks process termination.

Logical edge IDs and execution IDs correlate messages. Stale notifications must not overwrite terminal execution state or trigger duplicate notifications.

Telemetry

Hayata uses Telemetry to expose lifecycle events from agents, nodes, edges, and reducers. Applications can attach handlers to collect metrics, trace execution activity, or integrate with monitoring systems.

A stop event indicates a terminal outcome, not necessarily success. Where present, use the status metadata to distinguish completed, failed, cancelled, and timed-out executions.

Start events include system_time. Terminal events include duration, calculated with erlang:monotonic_time/0. Both use Erlang native time units; convert durations when exporting to systems that expect milliseconds.

Common metadata includes agent_id, node_id, execution_id, edge_id, status, and, for caught exceptions, class. Reducer events may also include expected_edges, result_count, error_count, and pending_edges. Available fields vary by event.

Hayata's telemetry metadata avoids including full results, exception reasons, or stacktraces. Applications should also review their own handlers and exporters to avoid exposing sensitive inputs, state, prompts, or outputs.

Attaching a handler

-module(my_hayata_telemetry).

-export([attach/0, detach/0, handle_event/4]).

-define(HANDLER_ID, {?MODULE, agent_invocations}).

attach() ->
    telemetry:attach_many(
        ?HANDLER_ID,
        [
            [hayata, agent, invocation, start],
            [hayata, agent, invocation, stop],
            [hayata, agent, invocation, exception]
        ],
        fun ?MODULE:handle_event/4,
        #{}
    ).

detach() ->
    telemetry:detach(?HANDLER_ID).

handle_event(Event, Measurements, Metadata, _Config) ->
    logger:info(
        "Hayata telemetry event",
        #{
            event => Event,
            measurements => Measurements,
            metadata => Metadata
        }
    ).

Call my_hayata_telemetry:attach/0 during application startup and my_hayata_telemetry:detach/0 when the handler is no longer needed. Keep handlers lightweight; forward expensive processing to a dedicated process or metrics backend.

Fault handling and process supervision

The edge watcher monitors the node independently. If the node dies, including through an untrappable kill, the watcher terminates the edge so it does not remain orphaned. Node processes are supervised and restarted after abnormal termination.

Edge tasks execute in linked worker processes, keeping the edge GenServer responsive while a task is running. If the edge terminates abnormally, its linked worker is terminated as well. However, terminating a process cannot undo external side effects that have already occurred.

If a task may have affected an external service before its result was durably saved, recovery reports task_execution_uncertain and does not automatically re-execute the task. Applications should use idempotency keys, external deduplication, or other application-level safeguards where retry safety matters.

Design principles

• Use OTP processes and monitoring for concurrency and failure detection.
• Keep process responsibilities and state ownership explicit.
• Treat logical IDs separately from process PIDs.
• Keep cancellation and recovery scoped to the affected execution.
• Persist task outcomes before treating them as completed.
• Reject uncertain execution states rather than silently repeating potentially non-idempotent work.
• Validate behavior incrementally with EUnit and explicit VM-restart checks.

Retrieval-Augmented Generation (RAG)

Hayata provides a provider-agnostic retrieval interface through the hayata_rag module. It allows nodes to retrieve relevant documents from an application-defined retrieval provider without coupling the framework to a particular vector database, embedding service, or search engine.

Hayata does not implement document indexing, chunking, embedding generation, or vector storage. Applications integrate their own retrieval infrastructure and provide a retriever function that implements the Hayata contract.

Configuring a retriever

Configure RAG using the rag_config key in the agent configuration. The retrieve entry accepts either an anonymous function with arity two or an MFA tuple pointing to an exported function with arity two.

The retriever receives a query and an options map. It must return {ok, #{documents := Documents}} or {error, Reason}.

Retriever = fun(Query, Options) ->
    %% Call your own retrieval infrastructure.
    my_retriever:search(Query, Options)
end,

Config = #{
    id => knowledge_agent,
    state => #{},
    nodes => Nodes,
    decision_policy => DecisionPolicy,
    rag_config => #{
        retrieve => Retriever
    }
}.

An MFA retriever can also be configured:

rag_config => #{
    retrieve => {my_retriever, search}
}

The provider is responsible for translating its native response into the normalized Hayata document contract.

Normalized document contract

Every retrieved document must contain:

A valid retriever response looks like this:

{ok, #{
    documents => [
        #{
            id => <<"chunk-001">>,
            content => <<"Erlang uses lightweight processes.">>,
            metadata => #{
                source => <<"erlang-guide.pdf">>,
                page => 12
            },
            score => 0.95
        }
    ]
}}

The score is optional because retrieval providers do not necessarily expose comparable relevance scores.

The hayata_rag module validates the retriever configuration and normalizes its result contract. Invalid responses are rejected, while provider errors are propagated.

Using RAG directly from a node

When rag_config is configured, Hayata injects a hayata_rag function into every node execution context.

The function accepts a binary query and an options map:

Rag(Query, Options)

It returns either a normalized document result or an error.

RetrieveNode = fun(Context) ->
    Rag = maps:get(hayata_rag, Context),

    case Rag(<<"Erlang process model">>, #{limit => 5}) of
        {ok, #{documents := Documents}} ->
            #{
                retrieved_documents => Documents
            };

        {error, Reason} ->
            #{
                retrieval_error => Reason
            }
    end
end.

RAG retrieval is independent of LLM configuration. An agent can retrieve and process documents without configuring an LLM provider.

Using RAG through LLM tool calling

RAG can also be exposed as a tool to an LLM through hayata_tool_loop. This allows the model to decide when retrieval is necessary.

The tool handler delegates to the same injected hayata_rag function. There is no separate retrieval implementation or RAG-specific tool runtime.

RagNode = fun(Context) ->
    Rag = maps:get(hayata_rag, Context),
    ToolLoop = maps:get(hayata_tool_loop, Context),

    SearchTool = #{
        name => <<"search_knowledge">>,
        description =>
            <<"Searches the knowledge base for relevant documents.">>,
        parameters => #{
            type => object,
            properties => #{
                query => #{
                    type => string,
                    description => <<"The information to search for.">>
                }
            },
            required => [<<"query">>]
        },
        handler => fun(Arguments) ->
            Rag(
                maps:get(query, Arguments),
                #{limit => 5}
            )
        end
    },

    Request = #{
        messages => [
            #{
                role => user,
                content => [
                    #{
                        type => text,
                        text => <<"Explain the system architecture.">>
                    }
                ]
            }
        ]
    },

    case ToolLoop(Request, [SearchTool], 5) of
        {ok, Result} ->
            #{answer => maps:get(response, Result)};

        {error, Reason} ->
            #{error => Reason}
    end
end.

This example requires both rag_config and an LLM configuration in the agent. The hayata_tool_loop capability is available when an LLM is configured.

Design boundaries

Hayata deliberately keeps retrieval separate from generation:

• hayata_rag handles document retrieval and result validation.
• hayata_llm handles text generation.
• hayata_tool_loop coordinates LLM tool calling and execution.
• Applications own their retrieval infrastructure, indexing, embeddings, and document lifecycle.

This separation allows applications to use different retrieval technologies without requiring provider-specific adapters in the Hayata core.

LLM

Hayata's LLM integration is designed around a provider-neutral adapter contract. The shared contract defines the normalized request and response shapes, capability reporting, error codes, and optional streaming interface. Provider-specific API payloads and HTTP details stay inside each adapter.

The hayata_tool_loop coordinates model calls and tool execution independently of the provider: it validates normalized assistant responses, executes registered tools in order, adds tool results to the conversation history, and continues until the model returns a final response or the iteration limit is reached. Providers should return normalized tool calls; they should not execute Hayata tool handlers themselves.

Available adapters

• OpenAI — OpenAI Responses API adapter documentation. Supports normalized text generation, tool calling, and streaming.

Implementing another provider

Use the Adding an LLM provider guide as the implementation checklist. It covers the adapter contract, normalized errors, tool-call serialization, streaming guarantees, and the test coverage expected before a provider is listed here.

License

Apache-2.0. See LICENSE.md for details.