portlang Skill

Core Concept

portlang treats agent behavior as search through a conditioned space. You don't script loops—you declare the search space:

• Boundaries: What the agent cannot do (enforced by sandbox)
• Verifiers: What success looks like (deterministic pass/fail signals, not walls of output)
• Context budget: Hard token ceiling
• Environment: What the agent can observe

The runtime executes the search. Every run produces a trajectory (complete event log).

The instruction budget problem: Models follow roughly 150–200 instructions reliably. A large system prompt, many tool definitions, growing conversation history, and re-observations all compete for that budget. portlang keeps budget usage low by design: each field is scoped narrowly, context is bounded by a hard ceiling, and re-observations inject only current state rather than accumulating history.

Prerequisites

portlang currently only runs on apple devices. Before running portlang fields: