Research Phase

Bounded Rationality

AI agents that know their limits — and work brilliantly within them.

Herbert A. Simon (1916-2001) — Nobel laureate in Economics, Turing Award winner. His work on bounded rationality revolutionized how we think about decision-making under constraints. Nika applies these principles to AI agent orchestration.

The Problem with "Optimal" AI

Most AI frameworks assume unlimited resources: infinite context windows, unlimited API calls, perfect information. But production systems have real constraints:

Token budgets run out mid-reasoning
API rate limits throttle parallel agents
Context windows can't hold all relevant information
Latency constraints demand fast decisions

Simon's Solution: Satisficing

"Satisficing = Satisfy + Suffice"

Instead of searching for the best answer, find the first good enough answer. Stop when you hit an acceptable threshold, not when you've exhausted all options.

This isn't settling for mediocrity — it's recognizing that perfect is the enemy of shipped. In agentic workflows, a fast acceptable result beats a slow optimal one every time.

How Nika Implements Bounded Rationality

Satisficing Over Optimizing

Agents find "good enough" solutions quickly rather than exhaustively searching for optimal ones. First acceptable answer wins.

"Satisficing = Satisfy + Suffice"

Cognitive Budget Limits

Each agent has explicit token budgets, turn limits, and timeout constraints. No infinite loops, no runaway costs.

"Bounded by computational capacity"

Heuristic Decision Rules

Simple rules beat complex optimization under uncertainty. Agents use fast-and-frugal heuristics for routing and fallback.

"Use of heuristics to reduce complexity"

Explicit Uncertainty

Agents acknowledge what they don't know. Epistemic signals surface uncertainty rather than hiding it in confidence scores.

"Incomplete information is the norm"

Traditional vs Bounded AI Agents

Aspect	Traditional AI	Bounded Rationality
Goal	Find optimal solution	Find acceptable solution fast
Resources	Unlimited compute	Explicit budgets & limits
Uncertainty	Hidden in probabilities	Surfaced as signals
Failure	Retry until success	Fail fast, escalate early
Complexity	Handle all edge cases	Handle common cases well

In Practice: Budget Constraints

bounded-agent.nika.yaml

tasks:
  - id: analyze
    agent:
      prompt: "Analyze this codebase for security issues"
      model: claude-sonnet-4-5
      scopePreset: minimal      # Fresh context, no accumulation
      maxTurns: 10              # Hard limit: satisfice by turn 10
      timeout: 30000            # 30s max: good enough beats perfect
      tools: [Read, Grep]       # Limited tools = faster decisions

  - id: fallback
    agent:
      prompt: "Quick security scan, top 3 issues only"
      model: claude-haiku       # Cheaper, faster fallback
      maxTurns: 3               # Ultra-bounded for speed

flows:
  - source: analyze
    target: fallback
    condition: timeout          # If main agent times out, satisfice

Research Status

Bounded rationality principles inform Nika's design philosophy. We're actively researching:

Budget enforcement (maxTurns, timeout) — Implemented
Scope isolation for cognitive limits — Implemented
Adaptive satisficing thresholds — In Research
Heuristic-based routing decisions — In Research

All Concepts Graceful Degradation