Agent observability — from one contract

Everything on this page is derived statically at build time by @suluk/agents from a single x-suluk-agents declaration — no runtime, no LLM call. It is the same "contracts in, everything derived" idea that powers the rest of saasuluk, applied to agents: the pyramid, a hardening grade, token-budget and context-waste warnings, the pattern an agent's shape affords, its loadable resources, and its paid / human-gated tools.

The pyramid — a determinism gradient

Routes (no model) are the deterministic floor; each layer above composes lower layers + a model. Higher = less deterministic, more general.

LAYER 2 · top (most general)

supportCopilot A

1 skill(s) · 4 route(s) · 2 sub-agent(s)

context ≈ 641 tok / budget 12,000

LAYER 1

escalation A

1 skill(s) · 1 route(s) · 0 sub-agent(s)

context ≈ 419 tok

researcher A

1 skill(s) · 1 route(s) · 0 sub-agent(s)

context ≈ 420 tok

LAYER 0 · deterministic floor (routes → MCP tools)

searchOrders · getOrder · issueRefund · draftReply

What each agent's shape affords

Structural capability only — the runtime trajectory stays opaque by design (the contract never claims which pattern actually runs).

supportCopilot

evaluator-optimizer orchestrator-workers parallelization routing

researcher

single-step (no multi-step pattern)

escalation

single-step (no multi-step pattern)

Hardening grade · resources · paid + gated tools

Hardening

A supportCopilot · 4 finding(s) A researcher · 1 finding(s) A escalation · 1 finding(s)

Loadable resources (CF "Agent Skills")

refundPolicy · instructions replyStyle · reference

Paid + human-gated

issueRefund → x402 paidTool $0.03/call needs human approval ⚖ — "money leaves the account"

Read the API @suluk/agents on npm