diff --git a/CHANGELOG.md b/CHANGELOG.md
index e515800..4aeaf8a 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,7 @@ All notable changes to SwiftAgentKit will be documented in this file.
 ## Unreleased
 
 ### Added
+- **Autonomous mode.** `AgentConfig(autonomousMode:)` (default `false`) and a runtime `Agent.setAutonomousMode(_:)` / `ToolDispatcher.setAutonomousMode(_:)` let an app skip the `requiresConfirmation` gate so the agent runs confirmation-required tools without prompting `AgentCallbacks.onToolConfirmation`. Off by default (confirmation-gated / fail-closed unchanged); flip it on for full-autonomy runs. Regression coverage added (autonomous mode bypasses the gate with no handler present).
 - **Context management (ContextSift), opt-in.** A new `ContextManager` (in `Sources/SwiftAgentKit/Context/`) keeps the main user/assistant conversation intact while externalizing *completed* tool exchanges: each finished tool result becomes a compact one-line receipt in a tool ledger (folded into the system prompt), and its full output is offloaded to an `ArtifactStore` (default `InMemoryArtifactStore`) so it can be retrieved losslessly on demand rather than resent every turn. The single in-flight ("active") tool exchange is preserved in full (bounded by `maxActiveResultChars`), so the model always has what it needs to act next. Two retrieval tools — `artifact_read` (paged) and `artifact_search` (line-level substring) — are auto-registered when a `ContextManager` is set, letting the model pull back any offloaded output by its `artifact-…` id. Enable per-agent via `AgentConfig(contextManager:)`; leaving it `nil` preserves the previous behavior exactly (`makeLLMRequest` is unchanged when unset). Ports the Python ContextSift approach (compact receipts + artifact store + on-demand retrieval) into the framework. Regression coverage added (completed exchanges → ledger + artifact and dropped from context; active exchange kept; large results spilled and round-tripped via `artifact_read`; artifact store read/search).
 - Guard against overlapping `run(_:)` calls on the same `Agent` instance with `AgentError.runInProgress`.
 - Regression coverage for same-instance concurrent run rejection.
diff --git a/Sources/SwiftAgentKit/Core/Agent.swift b/Sources/SwiftAgentKit/Core/Agent.swift
index 72885c3..4cf21f2 100644
--- a/Sources/SwiftAgentKit/Core/Agent.swift
+++ b/Sources/SwiftAgentKit/Core/Agent.swift
@@ -68,6 +68,12 @@ public struct AgentConfig: Sendable {
     /// agent uses its normal trim-based context handling.
     public var contextManager: ContextManager?
 
+    /// When `true`, tools marked `requiresConfirmation` run WITHOUT prompting via
+    /// `AgentCallbacks.onToolConfirmation` — the agent has full autonomy. Default
+    /// `false` (confirmation-gated). Can also be flipped at runtime with
+    /// `agent.setAutonomousMode(_:)`.
+    public var autonomousMode: Bool
+
     public init(
         provider: any LLMProvider,
         model: String? = nil,
@@ -82,7 +88,8 @@ public struct AgentConfig: Sendable {
         enableRepairRetry: Bool = true,
         enablePlanContinuation: Bool = true,
         tools: [any AgentTool] = [],
-        contextManager: ContextManager? = nil
+        contextManager: ContextManager? = nil,
+        autonomousMode: Bool = false
     ) {
         self.provider = provider
         self.model = model
@@ -98,6 +105,7 @@ public struct AgentConfig: Sendable {
         self.enablePlanContinuation = enablePlanContinuation
         self.tools = tools
         self.contextManager = contextManager
+        self.autonomousMode = autonomousMode
     }
 }
 
@@ -263,10 +271,21 @@ public final class Agent: @unchecked Sendable {
                 register(tool)
             }
         }
+
+        // Apply autonomous mode (skips the confirmation gate) if configured.
+        if config.autonomousMode {
+            setAutonomousMode(true)
+        }
     }
 
     // MARK: - Tools
 
+    /// Enable/disable autonomous mode at runtime. When `true`, tools marked
+    /// `requiresConfirmation` run without prompting `onToolConfirmation`.
+    public func setAutonomousMode(_ enabled: Bool) {
+        trackRegistrationTask(Task { await dispatcher.setAutonomousMode(enabled) })
+    }
+
     /// Register a tool.
     public func register(_ tool: any AgentTool) {
         trackRegistrationTask(Task { await tools.register(tool) })
diff --git a/Sources/SwiftAgentKit/Tools/ToolDispatcher.swift b/Sources/SwiftAgentKit/Tools/ToolDispatcher.swift
index 4e0eec1..5b4f4c2 100644
--- a/Sources/SwiftAgentKit/Tools/ToolDispatcher.swift
+++ b/Sources/SwiftAgentKit/Tools/ToolDispatcher.swift
@@ -45,6 +45,11 @@ public actor ToolDispatcher {
         contextFields = context
     }
 
+    /// Enable/disable autonomous mode (skips the `requiresConfirmation` gate).
+    public func setAutonomousMode(_ enabled: Bool) {
+        autonomousMode = enabled
+    }
+
     // MARK: - Dispatch
 
     /// Dispatch a batch of tool calls — optionally in parallel.
diff --git a/Tests/SwiftAgentKitTests/SwiftAgentKitTests.swift b/Tests/SwiftAgentKitTests/SwiftAgentKitTests.swift
index 8c2c0bd..d00271f 100644
--- a/Tests/SwiftAgentKitTests/SwiftAgentKitTests.swift
+++ b/Tests/SwiftAgentKitTests/SwiftAgentKitTests.swift
@@ -292,6 +292,21 @@ struct DangerousTool: AgentTool {
     #expect(results[0].result.contains("not approved"))
 }
 
+@Test func testAutonomousModeBypassesConfirmationGate() async {
+    let registry = ToolRegistry()
+    await registry.register(DangerousTool())
+    let dispatcher = ToolDispatcher(registry: registry)
+    await dispatcher.setAutonomousMode(true)
+    let state = AgentState()
+
+    // No onToolConfirmation handler, but autonomous mode is on → runs anyway.
+    let call = AgentToolCall(name: "delete_everything")
+    let results = await dispatcher.dispatch(calls: [call], state: state, observer: nil)
+    #expect(results.count == 1)
+    #expect(results[0].isError == false)
+    #expect(results[0].result == "done")
+}
+
 // MARK: - Conversation/Memory Tests
 
 @Test func testConversationAppendAndRead() {