plato-instinct

18 Agent Instincts with Constraint Grammar

MUST, SHOULD, CANNOT, MAY — formal behavioral constraints for AI agents.

What?

plato-instinct provides a formal grammar for agent behavioral constraints. Instead of vague instructions ("be careful"), agents get typed, enforceable instincts:

MUST — Hard constraint. Violation is a critical failure.
SHOULD — Soft constraint. Violation is logged and scored.
CANNOT — Negative constraint. The agent is structurally unable to perform the action.
MAY — Optional capability. The agent can but isn't required to.

Quick Start

[dependencies]
plato-instinct = "0.1"

use plato_instinct::*;

let engine = InstinctEngine::new();

// Define instincts with formal severity
engine.add(Instinct::new(
    "exfiltrate_credentials",
    Severity::MUST_NOT,       // Critical — never leak secrets
    Enforcement::Hard,         // Structurally enforced
    "Never expose fleet credentials to external channels"
));

engine.add(Instinct::new(
    "push_every_30min",
    Severity::SHOULD,          // Soft — preferred but not critical
    Enforcement::Soft,         // Scored but not blocked
    "Push work to vessel repos every 30 minutes"
));

// Evaluate agent behavior against instincts
let result = engine.evaluate(&action);

Core API

Type	Description
`Instinct`	A named behavioral constraint with severity and enforcement level
`Severity`	`MUST`, `SHOULD`, `CANNOT`, `MAY` — RFC 2119-style constraint levels
`Enforcement`	`Hard` (blocks action) or `Soft` (scores and logs)
`Assertion`	A checked behavioral claim against an instinct
`State`	Current instinct satisfaction state for an agent
`Reflex`	Automatic response triggered by instinct evaluation
`Thresholds`	Configurable thresholds for soft enforcement scoring
`InstinctEngine`	The full engine managing all instincts for an agent

Design Philosophy

Agent behavioral constraints should be machine-checkable, not just human-readable. By encoding instincts as typed data structures with formal enforcement semantics, we get:

Verifiable compliance — InstinctEngine::audit() checks all instincts
Composable constraints — instincts can reference and override each other
Scored behavior — soft constraints produce measurable compliance scores
Emergent discipline — hard constraints are structurally impossible to violate

Part of PLATO

Part of the PLATO ecosystem — formal knowledge and behavioral constraints for AI agent fleets.

License

MIT — Cocapn Fleet