2nd/10_Wiki/Topics/Backend/E-component (Execution Loop).md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-e-component-execution-loop	E-component (Execution Loop)	10_Wiki/Topics	verified	self	Execution Loop E-Loop Agent Runtime Loop	none	A	0.9	applied	agent runtime llm architecture		2026-05-10	pending	language framework Python anthropic-sdk

E-component (Execution Loop)

매 한 줄

"매 LLM agent 의 heartbeat". 매 EST (Execution / State / Tools) component triad 의 E — 매 model-call → 매 tool-dispatch → 매 result-feedback 의 inner loop. 매 Claude Agent SDK / OpenAI Assistants / LangGraph 매 same primitive.

매 핵심

매 mechanism

1. Send messages + tool definitions to LLM.
2. LLM 매 returns text + (optional) tool_use blocks.
3. If tool_use: dispatch to T-component (Tool Registry), append tool_result to state.
4. Loop until 매 stop_reason == "end_turn" or max_iterations.

매 components

• E (this): orchestrator — message-pump.
• S (State Store): conversation history, scratch state.
• T (Tool Registry): handler dispatch, schema validation.

매 응용

1. Code agents (Claude Code, Cursor, Devin).
2. Research agents (Perplexity, Deep Research).
3. Workflow automation.

💻 패턴

Minimal execution loop (Anthropic SDK 2026)

from anthropic import Anthropic
client = Anthropic()

def run(messages, tools, dispatch, max_iters=20):
    for _ in range(max_iters):
        resp = client.messages.create(
            model="claude-opus-4-7",
            max_tokens=4096,
            tools=tools,
            messages=messages,
        )
        messages.append({"role": "assistant", "content": resp.content})
        if resp.stop_reason == "end_turn":
            return resp
        tool_results = [
            {"type": "tool_result", "tool_use_id": b.id, "content": dispatch(b.name, b.input)}
            for b in resp.content if b.type == "tool_use"
        ]
        messages.append({"role": "user", "content": tool_results})
    raise RuntimeError("max iterations exceeded")

Streaming variant

with client.messages.stream(model="claude-opus-4-7", messages=messages, tools=tools) as stream:
    for event in stream:
        if event.type == "content_block_delta":
            print(event.delta.text, end="", flush=True)
    final = stream.get_final_message()

Tool dispatch (T-component plug-in)

TOOLS = {
    "read_file": lambda input: open(input["path"]).read(),
    "list_dir":  lambda input: os.listdir(input["path"]),
}
def dispatch(name, input):
    try: return TOOLS[name](input)
    except Exception as e: return f"Error: {e}"

Stop conditions

STOP = {"end_turn", "stop_sequence", "max_tokens"}
if resp.stop_reason in STOP: break

Prompt caching the system + tools

resp = client.messages.create(
    model="claude-opus-4-7",
    system=[{"type": "text", "text": SYSTEM, "cache_control": {"type": "ephemeral"}}],
    tools=[{**t, "cache_control": {"type": "ephemeral"}} for t in tools],
    messages=messages,
)

Budget guard

total_tokens = 0
while True:
    resp = client.messages.create(...)
    total_tokens += resp.usage.input_tokens + resp.usage.output_tokens
    if total_tokens > BUDGET: raise BudgetExceeded()

매 결정 기준

상황	Approach
Single-turn task	Direct API call
Multi-tool task	E-loop
Long-running workflow	E-loop + checkpointing (S)

기본값: E-loop + prompt caching + budget guard + max-iter clamp.

🔗 Graph

• 부모: Agent-Architecture
• 변형: ReAct
• 응용: LangGraph
• Adjacent: S-component (State Store) · T-component (Tool Registry)

🤖 LLM 활용

언제: building agent runtime, multi-step tool-use task. 언제 X: pure single-shot prompt (no tools, no state).

❌ 안티패턴

• No max-iter cap: 매 infinite-loop 의 risk.
• No budget guard: 매 unbounded cost.
• Recreating system prompt per turn: cache miss → 매 5-10x cost.

🧪 검증 / 중복

• Verified (Anthropic Messages API docs, Claude Agent SDK).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — E-component FULL with SDK 2026 loop patterns