MCP HubMCP Hub
Volver a habilidades

forage-resources

pjt222
Actualizado 2 days ago
6 vistas
17
2
17
Ver en GitHub
Otroai

Acerca de

Esta habilidad aplica optimización de colonias de hormigas y teoría de forrajeo para buscar eficientemente en grandes espacios de solución, equilibrando la exploración de nuevas opciones con la explotación de las conocidas y buenas. Ayuda a desplegar exploradores, reforzar rutas exitosas, detectar rendimientos decrecientes y adaptar estrategias dinámicamente. Úsela cuando la búsqueda por fuerza bruta sea impracticable, necesite asignar recursos entre oportunidades inciertas o para diagnosticar convergencia prematura en óptimos locales.

Instalación rápida

Claude Code

Recomendado
Principal
npx skills add pjt222/agent-almanac -a claude-code
Comando PluginAlternativo
/plugin add https://github.com/pjt222/agent-almanac
Git CloneAlternativo
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/forage-resources

Copia y pega este comando en Claude Code para instalar esta habilidad

Documentación

Forage Resources

Apply foraging theory + ant colony opt → systematically search, evaluate, exploit distributed resources — balance exploration unknown vs exploitation known yields.

Use When

  • Search large solution space, brute-force impractical
  • Balance invest between explore new vs deepen known good
  • Optimize resource alloc across uncertain opportunities
  • Design search strategies distributed teams/automated agents
  • Diagnose premature convergence (stuck local optima) or perpetual wandering (never commit)
  • Complement coordinate-swarm w/ specific resource-discovery patterns

In

  • Required: Resource being sought (info, compute, talent, solutions, opportunities)
  • Required: Search space (size, structure, known features)
  • Optional: Current strategy + failure mode
  • Optional: N available scouts/searchers
  • Optional: Cost exploration vs cost exploitation failure
  • Optional: Time horizon (short-term exploitation vs long-term exploration)

Do

Step 1: Map Landscape

Characterize resource env → select strategy.

  1. Resource type + distribution:
    • Concentrated: cluster rich patches (talent in specific communities)
    • Distributed: spread evenly (bugs across codebase)
    • Ephemeral: appear + disappear (market opportunities)
    • Nested: rich patches contain sub-patches diff scales
  2. Information landscape:
    • How much known about locations before foraging?
    • Scouts share info w/ foragers? (see coordinate-swarm for signal design)
    • Static or changing while foraging?
  3. Cost structure:
    • Cost per scout deployed (time, compute, money)
    • Cost exploiting low-quality (opportunity cost)
    • Cost missing high-quality (regret)

→ Characterized landscape w/ distribution, info, cost. Determines foraging model.

If err: completely unknown → max exploration (all scouts, no exploit) for fixed budget → build initial map. Switch to model once character clear.

Step 2: Deploy Scouts w/ Trail Marking

Exploratory agents into search space + instructions mark what find.

  1. Allocate scout % (start 20-30% of available)
  2. Scout behavior:
    • Move through space randomized/systematic
    • Evaluate each location (quick not deep)
    • Mark discoveries w/ signal strength proportional to quality:
      • High quality → strong trail
      • Medium → moderate
      • Low → weak or no signal
    • Return info to collective (signal deposit, report, broadcast)
  3. Scout pattern:
    • Random walk: unknown, uniform landscapes
    • Levy flight: long jumps + local clustering — patchy
    • Systematic sweep: grid/spiral — bounded, well-defined
    • Biased random: lean toward similar previous finds — clustered

→ Scouts deployed, depositing signals proportional to quality. Initial map emerges from reports.

If err: nothing initial sweep → (a) scout % too low (increase 50%), (b) wrong pattern (random walk → Levy flight for patchy), (c) quality miscalibrated (lower detection threshold).

Step 3: Trail Reinforcement

Positive feedback loops amplify successful paths, let unsuccessful fade.

  1. Forager follows trail + finds good:
    • Reinforce signal (increase strength)
    • Reinforced → more foragers → more reinforcement → exploitation
  2. Forager follows trail + finds nothing:
    • No reinforce (trail decays naturally)
    • Weakening → fewer foragers → fades → exploration resumes
  3. Reinforcement params:
    • Deposit: proportional to quality
    • Decay rate: trails lose X%/time
    • Saturation cap: max strength (prevents runaway single path)
Trail Reinforcement Dynamics:
┌─────────────────────────────────────────────────────────────────────┐
│                                                                     │
│  Strong trail ──→ More foragers ──→ If good: reinforce ──→ EXPLOIT │
│       ↑                                                      │      │
│       │                              If bad: no reinforce    │      │
│       │                                     │                │      │
│       │                                     ↓                │      │
│  Decay ←── Weak trail ←── Fewer foragers ←── Trail fades    │      │
│       │                                                      │      │
│       ↓                                                      │      │
│  No trail ──→ Scouts explore ──→ New discovery ──→ New trail ↗      │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

→ Self-regulating loop: good attract, poor abandoned. Balance via trail dynamics.

If err: all converge single trail (premature convergence) → decay too slow or cap too high. Increase decay, lower cap, or random exploration mandates (10% ignore trails). Fade too fast → reduce decay.

Step 4: Diminishing Returns

Monitor yields → know when shift exploit back to explore.

  1. Track yield/effort each active site:
    • Increasing → healthy, continue
    • Flat → approach saturation, begin scouting alts
    • Decreasing → diminishing, reduce foragers, increase scouts
  2. Marginal value theorem:
    • Compare current yield vs avg across known sites
    • Current drops below avg → time to leave
    • Factor travel cost (switching to new)
  3. Trigger scouting waves:
    • Overall yield across all drops below threshold
    • Best-performing exploited longer than expected lifetime
    • Env change detected (new signals from unexplored areas)

→ Swarm shifts between exploit (known-good) + exploration (scouts dispersed), driven by monitoring not arbitrary schedules.

If err: stays depleted too long → marginal threshold too low or travel cost too high. Recalibrate via actual rates. Abandons good too early → threshold too sensitive, add smoothing window.

Step 5: Adapt Strategy

Select + switch strategies based on env feedback.

  1. Match to landscape:
    • Rich, clustered: commit heavy discovered patches (high exploit)
    • Sparse, scattered: high scout ratio (high explore)
    • Volatile, changing: short decay, frequent scouting waves (adaptive)
    • Competitive: faster reinforcement, pre-emptive marking (territorial)
  2. Monitor strategy-env mismatch:
    • High effort, low yield → too exploitative
    • High discovery, low follow-through → too exploratory
    • Oscillating yield → switching too aggressively
  3. Adaptive switching:
    • Rolling avg explore-to-exploit ratio
    • Ratio drifts too far from optimal (by landscape type) → nudge back
    • Gradual transitions — abrupt cause coordination chaos

→ System adapts balance to env, maintains effectiveness as conditions change.

If err: adaptation unstable (oscillating) → damping: require mismatch persist N time units before shift. No strategy works → reassess Step 1 landscape, distribution may be more complex than assumed.

Check

  • Landscape characterized (distribution, info, cost)
  • Scout % + pattern defined + deployed
  • Trail reinforcement loop functional (deposit, decay, saturation)
  • Diminishing returns triggers rebalance exploit → explore
  • Strategy-env match monitored + adaptive switching
  • System recovers landscape changes (new/depleted)

Traps

  • Premature convergence: All pile on first good find, ignore better. Cure: mandatory exploration %, trail saturation caps, decay.
  • Perpetual exploration: Scouts find new but swarm never commits. Cure: lower quality threshold for reinforcement, reduce scout %.
  • Ignore travel costs: Switching has cost. Constantly jumping similar-quality → waste travel > gain. Factor travel into marginal value.
  • Static strategy dynamic landscape: Optimized for yesterday fails tomorrow. Build adaptation into loop not afterthought.
  • Conflate scout + forager quality: Good scouts (broad, quick) + good foragers (deep, thorough) require diff skills. Don't force both roles.

  • coordinate-swarm — foundational coordination underpinning signal design
  • build-consensus — swarm must collectively agree which patches prioritize
  • scale-colony — scaling operations as landscape/swarm grows
  • assess-form — morphic for system current state, complementary to landscape
  • configure-alerting-rules — alerting applicable to diminishing returns
  • plan-capacity — capacity planning shares explore-exploit framing
  • forage-solutions — AI self-application variant; maps ant colony to single-agent solution exploration w/ scout hypotheses + trail reinforcement

Repositorio GitHub

pjt222/agent-almanac
Ruta: i18n/caveman-ultra/skills/forage-resources
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

Habilidades relacionadas

llamaguard

Otro

LlamaGuard es el modelo de Meta de 7-8B parámetros para moderar las entradas y salidas de LLM en seis categorías de seguridad como violencia y discurso de odio. Ofrece una precisión del 94-95% y puede implementarse usando vLLM, Hugging Face o Amazon SageMaker. Utiliza esta skill para integrar fácilmente filtrado de contenido y barreras de seguridad en tus aplicaciones de IA.

Ver habilidad

cost-optimization

Otro

Esta Skill de Claude ayuda a los desarrolladores a optimizar los costes en la nube mediante el ajuste de tamaño de recursos, estrategias de etiquetado y análisis de gastos. Proporciona un marco para reducir los gastos en la nube e implementar una gobernanza de costes en AWS, Azure y GCP. Úsala cuando necesites analizar los costes de infraestructura, ajustar el tamaño de los recursos o cumplir con restricciones presupuestarias.

Ver habilidad

quantizing-models-bitsandbytes

Otro

Esta habilidad cuantiza LLMs a precisión de 8 o 4 bits utilizando bitsandbytes, logrando una reducción de memoria del 50-75% con pérdida mínima de precisión. Es ideal para ejecutar modelos más grandes en memoria GPU limitada o para acelerar la inferencia, admitiendo formatos como INT8, NF4 y FP4. La habilidad se integra con HuggingFace Transformers y permite entrenamiento QLoRA y optimizadores de 8 bits.

Ver habilidad

dispatching-parallel-agents

Otro

Esta Skill de Claude despliega múltiples agentes para investigar y solucionar 3 o más problemas independientes de forma concurrente. Está diseñada para escenarios que involucran fallos no relacionados que pueden resolverse sin estado compartido o dependencias. Su capacidad principal es la resolución paralela de problemas, asignando un agente por cada dominio problemático independiente para maximizar la eficiencia.

Ver habilidad