Compel agents to keep acting until they explicitly finish
Forcing only the opening turn let a weak model take one token action and stop, so a kickoff hired one agent (or none) and went quiet. Now every turn is forced (tool_choice required) and the agent ends only by calling the new `finish` tool — so it must keep hiring, delegating, and messaging until the work is actually done. Finishing with zero actions, or returning no tool call at all, fails the run loudly instead of faking success. Prompts tell each agent it must call a tool every turn and finish only when done. Tests drive weak-model stubs that act then finish, proving a full team forms rather than a single gesture. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
1 parent
ef9ee463c9
commit
f7543e506d
7 files changed
+72
-50
No files matched your search
@@ -19,7 +19,8 @@ function fakeProvider(script: ScriptedTurn[]) {
|
||||
let index = 0;
|
||||
return {
|
||||
runToolTurn: async (): Promise<AgentTurnResult & { requestId: string }> => {
|
||||
const turn = script[index] ?? { text: "done", toolCalls: [] };
|
||||
// A run never ends on prose now; an exhausted script ends by calling finish.
|
||||
const turn = script[index] ?? { text: "done", toolCalls: [call("finish", {})] };
|
||||
index += 1;
|
||||
return { text: turn.text, toolCalls: turn.toolCalls, rawResponseJson: "{}", requestId: `req_${index}` };
|
||||
},
|
||||
@@ -33,14 +34,18 @@ function call(name: string, args: Record<string, unknown>): ToolCall {
|
||||
/** Mimics a weak model: it rambles in prose and skips tools unless the backend
|
||||
* compels a tool call. With forcing it posts a kickoff message, then stops. */
|
||||
function weakModelProvider() {
|
||||
let forced = false;
|
||||
let acted = false;
|
||||
return {
|
||||
runToolTurn: async (input: { toolChoice?: "auto" | "required" }): Promise<AgentTurnResult & { requestId: string }> => {
|
||||
if (input.toolChoice === "required" && !forced) {
|
||||
forced = true;
|
||||
return { text: "Acting.", toolCalls: [call("send_message", { channel: "general", body: "Kicking off the company." })], rawResponseJson: "{}", requestId: "req_forced" };
|
||||
// Left to itself this model rambles (no tools); compelled, it posts, then ends.
|
||||
if (input.toolChoice !== "required") {
|
||||
return { text: "Let me think about this some more.", toolCalls: [], rawResponseJson: "{}", requestId: "req_prose" };
|
||||
}
|
||||
return { text: "Let me think about this some more.", toolCalls: [], rawResponseJson: "{}", requestId: "req_prose" };
|
||||
if (!acted) {
|
||||
acted = true;
|
||||
return { text: "Acting.", toolCalls: [call("send_message", { channel: "general", body: "Kicking off the company." })], rawResponseJson: "{}", requestId: "req_act" };
|
||||
}
|
||||
return { text: "Done.", toolCalls: [call("finish", {})], rawResponseJson: "{}", requestId: "req_finish" };
|
||||
},
|
||||
};
|
||||
}
|
||||
@@ -92,7 +97,7 @@ describe("agent run loop (real tool dispatch)", () => {
|
||||
call("write_deliverable", { path: "plan/operating-plan.md", content: "# Operating Plan\n\n1. Validate demand.\n2. Launch booking site.\n" }),
|
||||
call("send_message", { channel: "general", body: "Team staffed and the operating plan is live. Booking site is next." }),
|
||||
] },
|
||||
{ text: "Team staffed, work delegated, plan written.", toolCalls: [] },
|
||||
{ text: "Team staffed, work delegated, plan written.", toolCalls: [call("finish", {})] },
|
||||
];
|
||||
|
||||
let toolResultNotifications = 0;
|
||||
@@ -181,7 +186,7 @@ describe("agent run loop (real tool dispatch)", () => {
|
||||
};
|
||||
const script: ScriptedTurn[] = [
|
||||
{ text: "escape", toolCalls: [call("write_deliverable", { path: "../escape.txt", content: "x" })] },
|
||||
{ text: "done", toolCalls: [] },
|
||||
{ text: "done", toolCalls: [call("finish", {})] },
|
||||
];
|
||||
const result = await runAgentLoop(
|
||||
{ providers: fakeProvider(script), tools: { db: connection.db, organization, chat: new ChatService(connection.db) } },
|
||||
@@ -213,7 +218,7 @@ describe("agent run loop (real tool dispatch)", () => {
|
||||
await expect(runAgentLoop(
|
||||
{ providers: fakeProvider([{ text: "Here is my plan: first I will think about it.", toolCalls: [] }]), tools: { db: connection.db, organization, chat: new ChatService(connection.db) } },
|
||||
{ orgId, providerId, workspaceId: null, instructions: "i", initialUserTurn: "u", toolContext: ctx },
|
||||
)).rejects.toThrow(/no action on its first turn/);
|
||||
)).rejects.toThrow(/no tool call on a forced turn/);
|
||||
connection.close();
|
||||
});
|
||||
});
|
||||
Reference in new issue
Block a user