ACT - Action-Only Agents

Theory

Paper

Original paper

Function Calling / Tool Use in LLMs Based on OpenAI Function Calling (2023) and Anthropic Tool Use ACT agents are the simplest form of tool-using agents and rely on the native function-calling capabilities of modern LLMs such as GPT-4, Claude, and Mistral.

Concept

ACT (Action-Only) executes actions directly without explicit reasoning traces. The LLM immediately selects the appropriate tool based on the request and executes it. Reasoning happens implicitly in the model weights.

Architecture

flowchart LR
    query([Query]) --> service[AgentChatService]
    service --> act[ACT Mode]
    act --> action[ACTION]
    action --> tool[[Tool Execute]]
    tool --> obs[/OBSERVATION/]
    obs --> decide{respond?}
    decide -->|No| act
    decide -->|Yes| answer([Answer])

    style query fill:#e8c87a,stroke:#d4a84b,color:#000
    style service fill:#a8c5e2,stroke:#7ba3c9,color:#000
    style act fill:#D1BC8A,stroke:#b8a06a,color:#000
    style tool fill:#b0ca97,stroke:#8fb077,color:#000
    style obs fill:#88c4c8,stroke:#5fa8ad,color:#000
    style decide fill:#9e9e9e,stroke:#757575,color:#fff
    style answer fill:#98d4bb,stroke:#6bbf9a,color:#000

Simple loop: Query → ACT → Action → Tool → Observation → (repeat or answer)

Core Concept

Action → Observation → Action → ... → Response

Compared to ReACT:

• No THOUGHT step - LLM chooses the action directly
• Implicit reasoning - hidden inside model weights
• Faster - fewer tokens per iteration
• Less transparent - no explanation of decisions

Comparison

Aspect	ACT	ReACT
Reasoning	Implicit	Explicit
Tokens per iteration	~50-100	~150-300
Interpretability	Low	High
Speed	Fast	Medium
Complex tasks	Limited	Good

When to Use ACT

• Simple lookup tasks ("What is the phone number?")
• Fast response times are important
• Transparency is not required
• Users do not expect explanations

---

Example

Query

"What are the opening hours?"

ACT Trace

[Iteration 1]
ACTION: rag_search("opening hours")
OBSERVATION:
  [1] Opening hours: Mon-Fri 9-18, Sat 10-14
  [2] Contact: You can reach us by phone...

[Iteration 2]
ACTION: respond("Opening hours are Monday to Friday
                 from 9 to 18 and Saturday from 10 to 14. [1]")

Response

Opening hours are Monday to Friday from 9 to 18
and Saturday from 10 to 14. [1]

Sources:
[1] Opening hours - Contact

---

Implementation in LLARS

Status: Production

ACT is fully implemented and in production use.

Architecture

flowchart TB
    subgraph main[chat_act Loop]
        direction TB
        start([Start]) --> gen[Generate ACTION]
        gen --> parse[Parse ACTION]
        parse --> check{respond?}
        check -->|No| exec[Execute Tool]
        exec --> hist[Add Observation]
        hist --> gen
    end
    check -->|Yes| done([Finalize])

    style start fill:#e8c87a,stroke:#d4a84b
    style gen fill:#a8c5e2,stroke:#7ba3c9,color:#000
    style parse fill:#D1BC8A,stroke:#b8a06a,color:#000
    style check fill:#9e9e9e,stroke:#757575,color:#fff
    style exec fill:#b0ca97,stroke:#8fb077,color:#000
    style hist fill:#88c4c8,stroke:#5fa8ad,color:#000
    style done fill:#98d4bb,stroke:#6bbf9a,color:#000
    style main fill:#f5f5f5,stroke:#D1BC8A

System Prompt

# DEFAULT_ACT_SYSTEM_PROMPT (db/models/chatbot.py)
"""
Du hast Zugriff auf folgende Tools:
- rag_search(query): Semantische Suche in den Dokumenten
- lexical_search(query): Woertliche Suche in den Dokumenten
- respond(answer): Finale Antwort geben

Nutze web_search nur, wenn es fuer diesen Bot aktiviert ist und in der Tool-Liste angegeben wird.
Nutze Suchbegriffe aus der aktuellen Nutzerfrage oder dem Verlauf.
Wenn die Frage ohne Kontext unklar ist, stelle eine Rueckfrage mit respond.
Schreibe keine [TOOL_CALLS]-Marker oder JSON-Toolcalls, sondern nur das ACTION-Format.

Fuehre die passende Aktion aus, um die Frage zu beantworten.
Format: ACTION: tool_name(parameter)
"""

Additionally: - chatbot.system_prompt is prefixed as base context. - build_tool_availability_prompt() adds the actually enabled tools dynamically. - {PROJECT_URL} placeholders are replaced before use.

Files

File	Function
app/services/chatbot/agent_chat_service.py	Routing to ACT/ReAct/ReflAct
app/services/chatbot/agent_modes/mode_act.py	chat_act() loop + streaming
app/services/chatbot/agent_tools.py	Tool execution + confidence checks
app/services/chatbot/agent_prompts.py	Prompt builder (ACT prompt + tool list)
app/services/chatbot/agent_parsers.py	ACTION parser
app/db/models/chatbot.py	DEFAULT_ACT_SYSTEM_PROMPT + prompt settings

Code Snippet

# mode_act.py - chat_act()
for iteration in range(max_iterations):
    yield {"status": "iteration", "iteration": iteration + 1, "max": max_iterations}

    # Generate ACTION (streaming)
    yield {"status": "getting_action", "iteration": iteration + 1}
    action_text = "..."

    # Parse ACTION
    action, param = parse_action(action_text)
    yield {"status": "action", "action": action, "param": param, "iteration": iteration + 1}

    if action == "respond":
        yield {"status": "final_answer"}
        ...
        return

    # Execute tool
    result, sources = service._tool_executor.execute_tool(action, param, message, enabled_tools)
    yield {"status": "observation", "result_preview": result[:200], "iteration": iteration + 1}

Configuration

# ChatbotPromptSettings
agent_mode: str = "act"
task_type: str = "lookup" | "multihop"
agent_max_iterations: int = 5

tools_enabled: List[str] = ["rag_search", "lexical_search", "respond"]
web_search_enabled: bool = False
web_search_max_results: int = 5

tavily_api_key: Optional[str] = "..."  # only if web_search_enabled
act_system_prompt: str = "..."         # custom prompt (optional)

Tools

Tool	Function	Return
rag_search	Semantic search	Hits + relevance + sources
lexical_search	Keyword search	Hits + sources
web_search	Tavily web search (optional)	Web results + URLs
respond	Final response	Ends loop

Adaptive Iteration (High Confidence)

If the search returns high-confidence results, ACT exits early and generates a final answer immediately. Confidence is derived from source relevance scores (check_high_confidence).

---

Events (WebSocket)

# Streaming Events (excerpt)
yield {"status": "starting", "mode": "act"}
yield {"status": "iteration", "iteration": 1, "max": 5}
yield {"status": "getting_action", "iteration": 1}
yield {"status": "action_delta", "delta": "...", "iteration": 1}
yield {"status": "action", "action": "rag_search", "param": "...", "iteration": 1}
yield {"status": "observation_delta", "delta": "...", "iteration": 1}
yield {"status": "observation", "result_preview": "...", "iteration": 1}
yield {"status": "adaptive_iteration", "iteration": 1, "reason": "high_confidence"}
yield {"status": "adaptive_response", "reason": "high_confidence_results"}
yield {"status": "max_iterations_reached"}
yield {"status": "final_answer"}
yield {"delta": "..."}
yield {"done": True, "full_response": "...", "sources": [...]} 

Logs

[AgentChatService] ACT adaptive iteration: high confidence on iteration 2

Metrics

Stored in chatbot_messages.agent_trace:

• Actions and observations
• Number of iterations
• Adaptive exit (if triggered)

Additionally, chatbot_messages.stream_metadata contains:

• mode
• iterations
• sources_count
• adaptive_exit (optional)

---

See Also

• ReAct Agents
• ReflAct Agents
• RAG