2nd/10_Wiki/Topics/Backend/goal.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-goal	goal (Goal Definition in Software & Agents)	10_Wiki/Topics	verified	self	Goal Specification Objective Function Agent Goal	none	A	0.85	applied	goal-setting planning agents okr project-management		2026-05-10	pending	language framework Markdown / Python convention + agent planning libs

goal (Goal Definition in Software & Agents)

매 한 줄

"매 'goal' 매 software-eng / agent context 의 매 measurable success criterion + termination condition". OKR 의 KR, RL 의 reward, agent 의 stop condition, project brief 의 north star — 매 모든 context 에서 매 same shape: state X must hold. 2026 LLM agent 시대 매 'goal' 의 설계 가 prompt engineering 의 most-leveraged part.

매 핵심

매 Goal anatomy

1. Predicate: state 가 hold 의 boolean function.
2. Metric: 진행도 의 measurable scalar.
3. Deadline: time bound.
4. Constraints: 매 do-not-violate (cost, safety).
5. Owner: 매 accountable entity.

매 Goal 의 levels

• Strategic (북극성, 분기): "$10M ARR by EoY".
• Tactical (epic, 매 sprint): "ship 3 enterprise features".
• Operational (PR, task): "reduce P95 latency from 800ms to 300ms".
• Agent-step (1 turn): "extract email from this PDF".

매 SMART 의 modern 개정

• Specific, Measurable, Aligned (was Achievable), Relevant, Time-boxed, Verifiable (new) — 매 LLM 의 self-check 가능.

매 응용

1. Project brief / _brief.md 의 "Why" + "Now" 섹션.
2. Agent system prompt 의 stop condition.
3. RL reward shaping.
4. PR description ("This PR achieves: ___").

💻 패턴

Goal struct (TypeScript)

interface Goal {
  id: string;
  description: string;
  predicate: () => Promise<boolean>;
  metric?: () => Promise<number>;       // higher = closer
  deadline?: Date;
  constraints: Constraint[];
  owner: string;
  parent?: string;                       // hierarchy
}

interface Constraint {
  type: 'budget' | 'latency' | 'safety';
  check: () => Promise<boolean>;
}

Agent stop-condition

async def run_agent(goal: Goal, max_steps=20):
    for step in range(max_steps):
        if await goal.predicate():
            return {"status": "achieved", "steps": step}
        for c in goal.constraints:
            if not await c.check():
                return {"status": "violated", "constraint": c.type}
        action = await llm_plan(goal, history)
        history.append(await execute(action))
    return {"status": "exhausted", "steps": max_steps}

OKR YAML

# goals/2026-Q2.yaml
objective: "Make Acme the default CRM for 50-person SMBs"
key_results:
  - id: arr
    description: "Reach $2M ARR"
    metric: arr_usd
    target: 2_000_000
    current: 1_240_000
  - id: nps
    description: "NPS ≥ 50"
    metric: nps_score
    target: 50
    current: 38
  - id: churn
    description: "Monthly churn < 2%"
    metric: monthly_churn
    target: 0.02
    direction: minimize

LLM goal-prompt template

You are working toward this goal:
<goal>{description}</goal>

You must terminate when ALL of these are true:
{predicate_checklist}

You must NOT violate:
{constraints}

After each action, output `<self-check>...</self-check>` where you state
whether the goal predicate is now true and why.

Hierarchical decomposition (HTN-style)

def decompose(goal: Goal) -> list[Goal]:
    # 매 LLM 또는 rule-based
    if goal.id == "ship_feature_X":
        return [
            Goal("design_doc", ..., parent=goal.id),
            Goal("api_impl", ..., parent=goal.id),
            Goal("ui_impl", ..., parent=goal.id),
            Goal("docs_update", ..., parent=goal.id),
        ]
    return [goal]

Verifiable goal check

const goals: Goal[] = [{
  id: 'p95_latency',
  description: 'p95 < 300ms for /search',
  predicate: async () => {
    const r = await fetch('https://prom/api/v1/query?query=histogram_quantile(0.95,rate(http_dur_bucket{route="/search"}[5m]))');
    const v = +(await r.json()).data.result[0].value[1];
    return v < 0.3;
  },
  constraints: [],
  owner: 'eng@acme',
}];

매 결정 기준

상황	Approach
Company-level	OKR (1 obj, 3-5 KRs, quarterly)
Team sprint	Goal + acceptance criteria checklist
LLM agent run	predicate + max-step + cost budget
RL training	dense reward + sparse goal + early stopping
Personal dev	weekly review, 매 1-2 active goals only

기본값: 매 written goal + measurable predicate + deadline + 매 weekly check-in. 매 매 unmeasurable goal 의 X.

🔗 Graph

• 부모: Project-Management
• 변형: OKR
• 응용: _brief
• Adjacent: KPI

🤖 LLM 활용

언제: goal decomposition draft, predicate code generation, OKR phrasing. 언제 X: 매 strategic priority 의 결정 — 매 founder/leadership 의 own.

❌ 안티패턴

• Vague goal: "improve UX" — 매 unverifiable. Use metrics.
• Too many goals: > 5 active = 0 active. Force prioritization.
• No constraints: agent 가 매 cost/safety 의 무시 → catastrophic.
• Goal-metric mismatch: Goodhart's law — metric 만 game 됨. Pair with qualitative review.
• Set-and-forget: 매 check-in 없으면 매 3개월 wasted.

🧪 검증 / 중복

• Verified (Doerr "Measure What Matters", Anthropic agent design notes, RL textbooks).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — goal anatomy + agent stop condition