Skip to content

Capabilities

Complete guide to Tachikoma's features and capabilities.

Core Capabilities

Intent Classification & Routing

Automatic intent classification routes requests to optimal execution strategies based on task complexity.

  • Low Complexity — Direct response (1-2s)
  • Medium Complexity — Single skill (5-15s)
  • High Complexity — Skill chain (15-45s)
  • Very High Complexity — RLM orchestration (45-120s)

Learn more →

Context Management

Position-aware context loading optimizes token placement for maximum effectiveness.

  • U-shaped attention bias awareness
  • Critical info at start/end
  • Context compression at 70-80% utilization
  • Structured summaries

Learn more →

Skill Execution

Specialized skills handle specific task types with optimal tool usage.

  • 5 core skills (down from 11)
  • Dynamic skill loading
  • Model-aware operations
  • Verification loops for critical tasks

Learn more →

Skill Chains

Orchestrate multiple skills for complex workflows.

  • Sequential execution
  • State passing between skills
  • Error handling
  • Verification integration

Learn more →

PAUL Methodology

PAUL (Plan-Apply-Unify Loop) — Structured development framework.

  • PLAN — Define objectives and acceptance criteria
  • APPLY — Execute with verification
  • UNIFY — Close loop

Never skip UNIFY — this is the heartbeat that prevents drift.

Learn more →

Advanced Capabilities

Model-Aware Editing

Dynamic edit format selection optimized for specific LLM models.

  • str_replace — Claude, Mistral (exact string matching)
  • str_replace_fuzzy — Gemini (fuzzy whitespace)
  • apply_patch — GPT (diff format)
  • hashline — Grok, GLM (content-hash anchoring)

Learn more →

Subagents

Workers for large-context discovery and parallel tasks.

  • Codebase exploration
  • Parallel search execution
  • Long-running sessions

Learn more →

Verification Loops

Generator-Verifier-Reviser pattern for high-stakes implementations.

  • GENERATE — Initial solution
  • VERIFY — Explicit criteria checking
  • REVISE — Fix based on feedback
  • Up to 3 iterations

Use for: complex implementations, high-stakes fixes, correctness-critical tasks. Skip for: simple tasks (<50 lines), prototypes, well-understood patterns.

Learn more →

Research-Backed Features

Meta Orchestration

Self-programming agent generation engine for AI-centered agent development.

  • Self-generating agent topology — Create specialized subagents on-demand
  • Dynamic tool synthesis — Write custom tools at runtime
  • Hierarchical memory management — Graph-based knowledge persistence

Learn more →

Cost-Aware Routing

Match task complexity to execution strategy for optimal speed vs accuracy.

Research: "When Do Tools and Planning Help LLMs Think?" (arXiv:2601.02663)

Learn more →

Position-Aware Context

Optimize for tokens-per-task using U-shaped attention patterns.

Research: "Found in the Middle" (ACL 2024)

Learn more →

Model Harness

Edit format selection matters as much as model choice.

Source: Can.ac blog (Feb 2026)

Learn more →

Capability Matrix

CapabilityBest ForComplexityLatency
Direct ResponseSimple queries, <50 linesLow1-2s
Single SkillFocused tasks, one domainMedium5-15s
Skill ChainMulti-step workflowsHigh15-45s
Verification LoopHigh-stakes, correctness-criticalHigh+10-30s
SubagentLarge-context discovery, parallel tasksVery High45-120s
RLMMassive contexts, 10M+ tokensVery High2-5min

Decision Flow

User Request

Classify Intent

Confidence > 0.7?
    ├── NO → Ask for clarification
    ↓ YES
Context > 2000 tokens?
    ├── YES → Use RLM subagent
    ↓ NO
Task Complexity?
    ├── Simple → Direct response
    ├── Medium → Single skill
    ├── High → Skill chain
    └── Critical → Verification loop

Reflect on approach (freedom to question, flag issues)

Quick Reference

NeedUseLink
Understand routingIntent Classification
Manage contextContext Management
Execute tasksSkill Execution
Chain skillsSkill Chains
Structure workPAUL Methodology
Model-specific editsModel-Aware Editing
Large contextsSubagents
Verify correctnessVerification Loops

Next Steps

Released under the MIT License.