refactor(agent): remove the inter-tool delay

The 1.0s sleep between sequential tool calls has been present since the
initial commit with no documented rationale. It sleeps between local
tool executions — the next LLM request goes out only after the whole
batch — so it rate-limits nothing, and the parallel read-only path
already runs with no delay. Every multi-tool turn pays (N-1) seconds
of dead time. Remove the sleep, the internal tool_delay plumbing, and
dead test assignments. AIAgent.__init__ keeps tool_delay as a
deprecated no-op keyword for one release so existing programmatic
callers construct cleanly; passing it emits a DeprecationWarning.

agent/agent_init.py

@@ -455,8 +455,7 @@ def init_agent(
     command: str = None,
     args: list[str] | None = None,
     model: str = "",
-    max_iterations: int = 500,  # Default tool-calling iterations (shared with subagents)
-    tool_delay: float = 1.0,
+    max_iterations: int = 90,  # Default tool-calling iterations (shared with subagents)
     enabled_toolsets: List[str] = None,
     disabled_toolsets: List[str] = None,
     save_trajectories: bool = False,
@@ -529,8 +528,7 @@ def init_agent(
         requested_provider (str): Original provider identity before runtime canonicalization
         api_mode (str): API mode override: "chat_completions" or "codex_responses"
         model (str): Model name to use (default: "anthropic/claude-opus-4.6")
-        max_iterations (int): Maximum number of tool calling iterations (default: 500)
-        tool_delay (float): Delay between tool calls in seconds (default: 1.0)
+        max_iterations (int): Maximum number of tool calling iterations (default: 90)
         enabled_toolsets (List[str]): Only enable tools from these toolsets (optional)
         disabled_toolsets (List[str]): Disable tools from these toolsets (optional)
         save_trajectories (bool): Whether to save conversation trajectories to JSONL files (default: False)
@@ -576,7 +574,6 @@ def init_agent(
     # Shared iteration budget — parent creates, children inherit.
     # Consumed by every LLM turn across parent + all subagents.
     agent.iteration_budget = iteration_budget or IterationBudget(max_iterations)
-    agent.tool_delay = tool_delay
     agent.save_trajectories = save_trajectories
     agent.verbose_logging = verbose_logging
     agent.quiet_mode = quiet_mode

agent/tool_executor.py

@@ -1979,9 +1979,6 @@ def execute_tool_calls_sequential(agent, assistant_message, messages: list, effe
                     return
             break
 
-        if agent.tool_delay > 0 and i < len(assistant_message.tool_calls):
-            time.sleep(agent.tool_delay)
-
     # ── Per-turn aggregate budget enforcement ─────────────────────────
     num_tools_seq = len(assistant_message.tool_calls)
     if finalize and num_tools_seq > 0:

run_agent.py

@@ -45,6 +45,7 @@ import tempfile
 import time
 import threading
 import uuid
+import warnings
 from typing import List, Dict, Any, Optional, Callable
 # NOTE: `from openai import OpenAI` is deliberately NOT at module top — the
 # SDK pulls ~240 ms of imports. We expose `OpenAI` as a thin proxy object
@@ -439,8 +440,8 @@ class AIAgent:
         command: str = None,
         args: list[str] | None = None,
         model: str = "",
-        max_iterations: int = 500,  # Default tool-calling iterations (shared with subagents)
-        tool_delay: float = 1.0,
+        max_iterations: int = 90,  # Default tool-calling iterations (shared with subagents)
+        tool_delay: float = None,  # Deprecated: accepted for compatibility, ignored
         enabled_toolsets: List[str] = None,
         disabled_toolsets: List[str] = None,
         save_trajectories: bool = False,
@@ -504,6 +505,13 @@ class AIAgent:
         requested_provider: str = None,
     ):
         """Forwarder — see ``agent.agent_init.init_agent``."""
+        if tool_delay is not None:
+            warnings.warn(
+                "tool_delay is deprecated and ignored; sequential tool calls "
+                "no longer sleep between executions.",
+                DeprecationWarning,
+                stacklevel=2,
+            )
         from agent.agent_init import init_agent
         init_agent(
             self,
@@ -518,7 +526,6 @@ class AIAgent:
             args=args,
             model=model,
             max_iterations=max_iterations,
-            tool_delay=tool_delay,
             enabled_toolsets=enabled_toolsets,
             disabled_toolsets=disabled_toolsets,
             save_trajectories=save_trajectories,

tests/run_agent/test_1630_context_overflow_loop.py

@@ -38,7 +38,6 @@ class TestGeneric400Heuristic:
             a.client = MagicMock()
             a._cached_system_prompt = "You are helpful."
             a._use_prompt_caching = False
-            a.tool_delay = 0
             a.compression_enabled = False
             return a
 

tests/run_agent/test_413_compression.py

@@ -94,7 +94,6 @@ def agent():
         a.client = MagicMock()
         a._cached_system_prompt = "You are helpful."
         a._use_prompt_caching = False
-        a.tool_delay = 0
         # Default matches production (`compression.enabled` defaults to True).
         # Overflow-recovery tests below verify that 413 / context-overflow
         # errors DO trigger compression; the disabled-path behavior is

tests/run_agent/test_conversation_fallback_state.py

@@ -80,7 +80,6 @@ def test_substantive_tool_only_turn_invalidates_older_housekeeping_fallback():
 
     agent._cached_system_prompt = "You are helpful."
     agent._use_prompt_caching = False
-    agent.tool_delay = 0
     agent.compression_enabled = False
     agent.save_trajectories = False
     agent.valid_tool_names = {"todo", "web_search"}
@@ -147,7 +146,6 @@ def test_housekeeping_only_turn_still_sets_fallback():
 
     agent._cached_system_prompt = "You are helpful."
     agent._use_prompt_caching = False
-    agent.tool_delay = 0
     agent.compression_enabled = False
     agent.save_trajectories = False
     agent.valid_tool_names = {"memory"}

tests/run_agent/test_malformed_tool_arguments.py

@@ -34,7 +34,6 @@ def _make_agent() -> AIAgent:
             skip_memory=True,
         )
     agent.client = MagicMock()
-    agent.tool_delay = 0
     agent._flush_messages_to_session_db = MagicMock()
     return agent
 

tests/run_agent/test_nonretryable_error_html_summary.py

@@ -72,7 +72,6 @@ def _make_agent() -> AIAgent:
     a.client = MagicMock()
     a._cached_system_prompt = "You are helpful."
     a._use_prompt_caching = False
-    a.tool_delay = 0
     a.compression_enabled = False
     a.save_trajectories = False
     return a

tests/run_agent/test_partial_stream_finish_reason.py

@@ -333,7 +333,6 @@ def loop_agent():
         a.client = MagicMock()
         a._cached_system_prompt = "You are helpful."
         a._use_prompt_caching = False
-        a.tool_delay = 0
         a.compression_enabled = False
         a.save_trajectories = False
         return a

tests/run_agent/test_run_agent.py

@@ -978,6 +978,25 @@ class TestInit:
             assert agent.api_mode == "anthropic_messages"
             mock_anthropic.Anthropic.assert_called_once()
 
+    def test_tool_delay_kwarg_is_deprecated_noop(self):
+        """tool_delay stays accepted for compatibility but warns and is ignored."""
+        with (
+            patch("run_agent.get_tool_definitions", return_value=[]),
+            patch("run_agent.check_toolset_requirements", return_value={}),
+            patch("run_agent.OpenAI"),
+        ):
+            with pytest.warns(DeprecationWarning, match="tool_delay"):
+                a = AIAgent(
+                    api_key="test-key-1234567890",
+                    base_url="https://openrouter.ai/api/v1",
+                    tool_delay=0,
+                    quiet_mode=True,
+                    skip_context_files=True,
+                    skip_memory=True,
+                )
+            # The value is discarded — nothing downstream reads it anymore.
+            assert not hasattr(a, "tool_delay")
+
     def test_prompt_caching_claude_openrouter(self):
         """Claude model via OpenRouter should enable prompt caching."""
         with (
@@ -2397,6 +2416,22 @@ class TestExecuteToolCalls:
         assert messages[0]["role"] == "tool"
         assert "search result" in messages[0]["content"]
 
+    def test_sequential_tool_calls_run_without_delay(self, agent):
+        """Two sequential tool calls execute back-to-back with no sleep between them."""
+        tc1 = _mock_tool_call(name="web_search", arguments="{}", call_id="c1")
+        tc2 = _mock_tool_call(name="web_search", arguments="{}", call_id="c2")
+        mock_msg = _mock_assistant_msg(content="", tool_calls=[tc1, tc2])
+        messages = []
+        with (
+            patch("run_agent.handle_function_call", return_value="ok") as mock_hfc,
+            patch("agent.tool_executor.time.sleep") as mock_sleep,
+        ):
+            agent._execute_tool_calls_sequential(mock_msg, messages, "task-1")
+        assert mock_hfc.call_count == 2
+        mock_sleep.assert_not_called()
+        tool_results = [m for m in messages if m["role"] == "tool"]
+        assert [m["tool_call_id"] for m in tool_results] == ["c1", "c2"]
+
     def test_sequential_memory_remove_notifies_provider_with_tool_result(self, agent):
         old_text = "stale preference entry"
         tc = _mock_tool_call(
@@ -4419,7 +4454,6 @@ class TestRunConversation:
         """Common setup for run_conversation tests."""
         agent._cached_system_prompt = "You are helpful."
         agent._use_prompt_caching = False
-        agent.tool_delay = 0
         agent.compression_enabled = False
         agent.save_trajectories = False
 
@@ -6702,7 +6736,6 @@ class TestRetryExhaustion:
     def _setup_agent(self, agent):
         agent._cached_system_prompt = "You are helpful."
         agent._use_prompt_caching = False
-        agent.tool_delay = 0
         agent.compression_enabled = False
         agent.save_trajectories = False
 
@@ -8703,7 +8736,6 @@ class TestReasoningReplayForStrictProviders:
     def _setup_agent(self, agent):
         agent._cached_system_prompt = "You are helpful."
         agent._use_prompt_caching = False
-        agent.tool_delay = 0
         agent.compression_enabled = False
         agent.save_trajectories = False
 

tests/run_agent/test_tool_call_guardrail_runtime.py

@@ -54,7 +54,6 @@ def _make_agent(*tool_names: str, max_iterations: int = 10, config: dict | None
     agent.client = MagicMock()
     agent._cached_system_prompt = "You are helpful."
     agent._use_prompt_caching = False
-    agent.tool_delay = 0
     agent.compression_enabled = False
     agent.save_trajectories = False
     return agent

tests/run_agent/test_tool_call_incremental_persistence.py

@@ -74,7 +74,6 @@ def _make_agent():
     agent.client = MagicMock()
     agent._cached_system_prompt = "You are helpful."
     agent._use_prompt_caching = False
-    agent.tool_delay = 0
     agent.compression_enabled = False
     agent.save_trajectories = False
     return agent

tests/run_agent/test_turn_completion_explainer.py

@@ -52,7 +52,6 @@ def _make_agent(max_iterations: int = 10, config: dict | None = None) -> AIAgent
     agent.client = MagicMock()
     agent._cached_system_prompt = "You are helpful."
     agent._use_prompt_caching = False
-    agent.tool_delay = 0
     agent.compression_enabled = False
     agent.save_trajectories = False
     # No fallback chain so empty responses exhaust deterministically.