agentic

forge:agentic

Model selection by capability

Don't pick by name. Pick by what the task needs. Names change; capabilities don't.

Task	Capability needed	Why
Complex reasoning, code gen	MAX tier	Needs deep multi-step reasoning
High-volume classification, triage	FAST tier	Throughput over depth
Structured extraction	Any + tool use	Schema enforced by tool, not model
Long document analysis	STANDARD+ with large context	Context window is the constraint
Real-time / streaming response	FAST tier	Latency is the constraint
Adversarial verification	MAX tier	Needs strong self-critique

For local models: start at STANDARD. Drop to FAST if context window < 32k or inference is slow.

Prompt template (4 required fields)

## Role         specific grounded persona with domain expertise
## Task         exact outcome required, not vague instructions
## Constraints  hard limits, what NOT to do
## Output Format schema or example, never leave format ambiguous

Anti-patterns: "be helpful", "do your best", giant unstructured system prompts.

MCP tool rules

Rule	Detail
Idempotent	Safe to retry without side effects
Schema first	description + JSON Schema on every tool
Structured errors	{error, code}, not raw stack traces
Minimal scope	Only permissions the tool actually needs

Cost optimization

Technique	Saving
Prompt caching on stable system prompts	~90% on repeated calls
Batch API for non-real-time work	~50%
Capability routing: FAST for triage, MAX for depth	60-80%
Context compression on long sessions	Reduces input tokens

Multi-agent pattern

Orchestrator
├── Researcher  (read-only, Explore agent)
├── Implementer (write, isolated worktree)
└── Reviewer    (read-only, forge-reviewer)

Use context: fork for sub-agents that must not affect main context. Use isolation: worktree for parallel file-writing agents.