refactor: updated base agent to accept runnables, deprecated llmChain

langchain/src/agents/agent.ts

@@ -276,8 +276,14 @@ export interface AgentArgs {
  * include a variable called "agent_scratchpad" where the agent can put its
  * intermediary work.
  */
-export abstract class Agent extends BaseSingleActionAgent {
-  llmChain: LLMChain;
+export abstract class Agent<
+  RunInput extends ChainValues & {
+    agent_scratchpad?: string | BaseMessage[];
+    stop?: string[];
+  } = any,
+  RunOutput extends AgentAction | AgentFinish = any
+> extends BaseSingleActionAgent {
+  runnable: Runnable<RunInput, RunOutput>;
 
   outputParser: AgentActionOutputParser | undefined;
 
@@ -288,13 +294,23 @@ export abstract class Agent extends BaseSingleActionAgent {
   }
 
   get inputKeys(): string[] {
-    return this.llmChain.inputKeys.filter((k) => k !== "agent_scratchpad");
+    // eslint-disable-next-line no-instanceof/no-instanceof
+    if (this.runnable instanceof LLMChain) {
+      return this.runnable.inputKeys.filter((k) => k !== "agent_scratchpad");
+    }
+    return [];
   }
 
   constructor(input: AgentInput) {
     super(input);
 
-    this.llmChain = input.llmChain;
+    if (!input.runnable && !input.llmChain) {
+      throw new Error(
+        `Runnable and LLMChain are both missing, one is required.`
+      );
+    }
+
+    this.runnable = input.runnable;
     this._allowedTools = input.allowedTools;
     this.outputParser = input.outputParser;
   }
@@ -385,12 +401,12 @@ export abstract class Agent extends BaseSingleActionAgent {
 
   private async _plan(
     steps: AgentStep[],
-    inputs: ChainValues,
+    inputs: RunInput,
     suffix?: string,
     callbackManager?: CallbackManager
   ): Promise<AgentAction | AgentFinish> {
     const thoughts = await this.constructScratchPad(steps);
-    const newInputs: ChainValues = {
+    const newInputs: RunInput = {
       ...inputs,
       agent_scratchpad: suffix ? `${thoughts}${suffix}` : thoughts,
     };
@@ -399,15 +415,16 @@ export abstract class Agent extends BaseSingleActionAgent {
       newInputs.stop = this._stop();
     }
 
-    const output = await this.llmChain.predict(newInputs, callbackManager);
+    const output = await this.runnable.invoke(newInputs, callbackManager);
     console.log({
       newInputs,
       output,
     });
-    if (!this.outputParser) {
-      throw new Error("Output parser not set");
-    }
-    return this.outputParser.parse(output, callbackManager);
+    return output;
+    // if (!this.outputParser) {
+    //   throw new Error("Output parser not set");
+    // }
+    // return this.outputParser.parse(output, callbackManager);
   }
 
   /**
@@ -421,7 +438,7 @@ export abstract class Agent extends BaseSingleActionAgent {
    */
   plan(
     steps: AgentStep[],
-    inputs: ChainValues,
+    inputs: RunInput,
     callbackManager?: CallbackManager
   ): Promise<AgentAction | AgentFinish> {
     return this._plan(steps, inputs, undefined, callbackManager);
@@ -433,7 +450,7 @@ export abstract class Agent extends BaseSingleActionAgent {
   async returnStoppedResponse(
     earlyStoppingMethod: StoppingMethod,
     steps: AgentStep[],
-    inputs: ChainValues,
+    inputs: RunInput,
     callbackManager?: CallbackManager
   ): Promise<AgentFinish> {
     if (earlyStoppingMethod === "force") {

langchain/src/agents/chat/index.ts

@@ -161,7 +161,7 @@ export class ChatAgent extends Agent {
       args?.outputParser ?? ChatAgent.getDefaultOutputParser();
 
     return new ChatAgent({
-      llmChain: chain,
+      runnable: chain,
       outputParser,
       allowedTools: tools.map((t) => t.name),
     });

langchain/src/agents/chat_convo/index.ts

@@ -207,7 +207,7 @@ export class ChatConversationalAgent extends Agent {
       callbacks: args?.callbacks ?? args?.callbackManager,
     });
     return new ChatConversationalAgent({
-      llmChain: chain,
+      runnable: chain,
       outputParser,
       allowedTools: tools.map((t) => t.name),
     });

langchain/src/agents/executor.ts

@@ -23,7 +23,7 @@ import { Runnable } from "../schema/runnable/base.js";
  * properties specific to agent execution.
  */
 export interface AgentExecutorInput extends ChainInputs {
-  agent: BaseSingleActionAgent | BaseMultiActionAgent | Runnable;
+  agent: BaseSingleActionAgent | BaseMultiActionAgent;
   tools: this["agent"]["ToolType"][];
   returnIntermediateSteps?: boolean;
   maxIterations?: number;

langchain/src/agents/mrkl/index.ts

@@ -145,7 +145,7 @@ export class ZeroShotAgent extends Agent {
     });
 
     return new ZeroShotAgent({
-      llmChain: chain,
+      runnable: chain,
       allowedTools: tools.map((t) => t.name),
       outputParser,
     });

langchain/src/agents/openai/index.ts

@@ -192,7 +192,7 @@ export class OpenAIAgent extends Agent {
       callbacks: args?.callbacks,
     });
     return new OpenAIAgent({
-      llmChain: chain,
+      runnable: chain,
       allowedTools: tools.map((t) => t.name),
       tools,
     });
@@ -233,7 +233,7 @@ export class OpenAIAgent extends Agent {
     }
 
     // Split inputs between prompt and llm
-    const llm = this.llmChain.llm as ChatOpenAI;
+    const llm = this.runnable.llm as ChatOpenAI;
     const valuesForPrompt = { ...newInputs };
     const valuesForLLM: (typeof llm)["CallOptions"] = {
       tools: this.tools,

langchain/src/agents/structured_chat/index.ts

@@ -212,7 +212,7 @@ export class StructuredChatAgent extends Agent {
     });
 
     return new StructuredChatAgent({
-      llmChain: chain,
+      runnable: chain,
       outputParser,
       allowedTools: tools.map((t) => t.name),
     });

langchain/src/agents/tests/agent.int.test.ts

@@ -11,7 +11,6 @@ import { WebBrowser } from "../../tools/webbrowser.js";
 import { Tool } from "../../tools/base.js";
 import { ChatOpenAI } from "../../chat_models/openai.js";
 import { RunnableSequence } from "../../schema/runnable/base.js";
-import { RunnableAgent } from "../agent.js";
 
 test("Run agent from hub", async () => {
   const model = new OpenAI({ temperature: 0, modelName: "text-babbage-001" });
@@ -63,8 +62,13 @@ test("Pass runnable to agent executor", async () => {
     outputParser,
   ]);
 
+  const agent = new ZeroShotAgent({
+    runnable,
+    allowedTools: tools.map((t) => t.name),
+  });
+
   const executor = AgentExecutor.fromAgentAndTools({
-    agent: runnable,
+    agent,
     tools,
   });
   const res = await executor.invoke({

langchain/src/agents/types.ts

@@ -1,6 +1,11 @@
 import { LLMChain } from "../chains/llm_chain.js";
 import { SerializedLLMChain } from "../chains/serde.js";
-import { AgentAction, AgentFinish, ChainValues } from "../schema/index.js";
+import {
+  AgentAction,
+  AgentFinish,
+  BaseMessage,
+  ChainValues,
+} from "../schema/index.js";
 import { BaseOutputParser } from "../schema/output_parser.js";
 import { Runnable } from "../schema/runnable/base.js";
 
@@ -9,11 +14,19 @@ import { Runnable } from "../schema/runnable/base.js";
  * LLMChain instance, an optional output parser, and an optional list of
  * allowed tools.
  */
-export interface AgentInput {
-  llmChain: LLMChain;
+export type AgentInput<
+  RunInput extends ChainValues & {
+    agent_scratchpad?: string | BaseMessage[];
+    stop?: string[];
+  } = any,
+  RunOutput extends AgentAction | AgentFinish = any
+> = {
+  /** @deprecated - use runnable instead */
+  llmChain?: LLMChain;
+  runnable: Runnable<RunInput, RunOutput>;
   outputParser: AgentActionOutputParser | undefined;
   allowedTools?: string[];
-}
+};
 
 /**
  * Interface defining the input for creating an agent that uses runnables.

langchain/src/schema/runnable/base.ts

@@ -54,8 +54,6 @@ export abstract class Runnable<
 > extends Serializable {
   protected lc_runnable = true;
 
-  declare ToolType: StructuredTool;
-
   abstract invoke(
     input: RunInput,
     options?: Partial<CallOptions>