diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.md
index 2e8ffcd37..92cead391 100644
--- a/packages/coding-agent/CHANGELOG.md
+++ b/packages/coding-agent/CHANGELOG.md
@@ -22,6 +22,7 @@
 
 ### Fixed
 
+- Fixed concurrent `createAgentSession` calls with the default agent id failing initialization with `Agent "Main" was replaced during session initialization` — each in-process embedder (e.g. the edit benchmark runner) can now pass a private registry via the newly exported `AgentRegistry`, keeping every top-level session's "Main" out of the process-global roster race.
 - Fixed task tool blocks duplicating their per-agent progress rows into terminal scrollback on every update: live task frames now pin the transcript live region so mid-run rows are never recorded as frozen snapshots, and a detached background task freezes its progress the moment any of its rows commit to scrollback instead of mutating committed history.
 - Fixed Codex reset fireworks comparing different quota tiers or plans, preventing false celebrations when usage reports switch between Spark and base weekly limits.
 - Fixed Cursor ranged-read results losing the full file byte size after applying the requested window.
diff --git a/packages/coding-agent/src/sdk.ts b/packages/coding-agent/src/sdk.ts
index 16f9e2716..0b0078982 100644
--- a/packages/coding-agent/src/sdk.ts
+++ b/packages/coding-agent/src/sdk.ts
@@ -609,6 +609,10 @@ export type * from "./extensibility/extensions";
 export type { Skill } from "./extensibility/skills";
 export type { FileSlashCommand } from "./extensibility/slash-commands";
 export type { MCPManager, MCPServerConfig, MCPServerConnection, MCPToolsLoadResult } from "./mcp";
+// Agent registry: pass a private instance per `createAgentSession` when
+// embedding several concurrent top-level sessions in one process (the default
+// global registry admits only one "Main" per process generation).
+export { type AgentRef, AgentRegistry, MAIN_AGENT_ID } from "./registry/agent-registry";
 export type { Tool } from "./tools";
 export { buildDirectoryTree, buildWorkspaceTree, type DirectoryTree, type WorkspaceTree } from "./workspace-tree";
 
diff --git a/packages/typescript-edit-benchmark/src/in-process-client.ts b/packages/typescript-edit-benchmark/src/in-process-client.ts
index 39995efbf..99f17fde7 100644
--- a/packages/typescript-edit-benchmark/src/in-process-client.ts
+++ b/packages/typescript-edit-benchmark/src/in-process-client.ts
@@ -9,6 +9,7 @@ import type { AgentEvent, AgentMessage, ResolvedThinkingLevel, ThinkingLevel } f
 import type { Model, ToolExample } from "@oh-my-pi/pi-ai";
 import type { AgentSession, AgentSessionEvent, AuthStorage, SessionStats } from "@oh-my-pi/pi-coding-agent";
 import {
+	AgentRegistry,
 	type CreateAgentSessionResult,
 	createAgentSession,
 	discoverAuthStorage,
@@ -97,6 +98,11 @@ export class InProcessClient {
 			authStorage: shared?.authStorage,
 			modelRegistry: shared?.modelRegistry,
 			sessionManager: SessionManager.inMemory(this.#options.cwd),
+			// Benchmark tasks run many top-level sessions concurrently in one
+			// process. The global registry admits only one "Main" per process
+			// generation (later registrations replace earlier refs, which then
+			// fail session initialization), so each client gets its own registry.
+			agentRegistry: new AgentRegistry(),
 			systemPrompt: this.#options.appendSystemPrompt
 				? (defaultPrompt: string[]) => [...defaultPrompt, this.#options.appendSystemPrompt!]
 				: undefined,