Skip to content

README.md

Latest commit

 

History

History
132 lines (90 loc) · 4.2 KB

README.md

File metadata and controls

132 lines (90 loc) · 4.2 KB
Raw

Active Inference Agent

Overview

This project implements an Active Inference inspired autonomous agent with a safety-oriented control plane.

The agent does not execute tool actions directly from a single planner output. Instead, it:

  • decomposes tasks into subtasks
  • simulates and evaluates actions before execution
  • applies a parallel testing suite as an execution gate
  • records outcomes and anomalies for later review
  • uses evidence-grounded fallback behavior for research-style tasks

Current Highlights

  • Active Inference loop: planning, simulation, evaluation, execution, and learning
  • Parallel Testing Suite: policy gate, outcome simulation, CNS, narrative pressure testing, watchdogs, and oversight memory
  • Fast-path optimizations: low-risk tools skip expensive safety/judge overhead
  • Fail-fast behavior: repeated dead-end searches no longer loop indefinitely
  • Evidence-grounded answering: research tasks avoid confident summaries when external evidence is missing
  • Persistent memory: semantic memory with episodic logging

Key Safety Features

Counter-Narrative Simulation

The agent simulates a malicious reinterpretation of a proposed action before execution. This helps detect dual-use or Trojan-horse prompts where a harmful task is disguised as a helpful one.

Narrative Pressure Test

The agent also tests whether a proposed action becomes unsafe under rhetorical pressure such as:

  • authority
  • urgency
  • secrecy
  • euphemism
  • false benevolence

This is implemented in parallel_testing/narrative_pressure_test.py.

Execution Gate

All actions pass through parallel_testing/execution_gate.py.

For low-risk tools, the gate uses a lightweight fast path.

For higher-risk tools, the gate runs:

  • outcome prediction
  • adversarial reinterpretation
  • narrative pressure scoring
  • pragmatic judging

Runtime Behavior

The current runtime includes several optimizations:

  • heuristic routing for repetitive low-risk subtasks
  • optional post-run LLM judge
  • single-sample judge by default
  • lower LLM timeout defaults
  • extract_info fast path for structured memory-style inputs

These changes improve responsiveness without removing the core safety gate.

Evidence-Grounded Fallback

For research-like tasks, the agent now checks whether retrieved evidence is actually relevant before presenting a summary.

If retrieval fails or evidence is irrelevant, the agent uses a transparent fallback instead of pretending it verified the topic.

Main Files

  • agent_manager.py: main orchestrator and active inference control loop
  • free_energy.py: EFE computation engine
  • llm_interpreter.py: policy and DAG planning
  • llm_gateway.py: model gateway for local or remote LLM calls
  • toolgate.py: tool execution layer
  • parallel_testing/events.py: event bus
  • parallel_testing/execution_gate.py: action validation gate
  • parallel_testing/outcome_simulator.py: outcome and adversarial simulation
  • parallel_testing/parallel_judge.py: pre-execution judging
  • parallel_testing/expectation_checker.py: predicted vs actual divergence tracking
  • parallel_testing/watchdog.py: anomaly watcher
  • parallel_testing/oversight_memory.py: rejection and oversight audit log
  • parallel_testing/narrative_pressure_test.py: rhetorical manipulation safety simulation

Running the Agent

python main.py "Research active inference and summarize key findings"

Useful Configuration

Environment variables:

EFE_THRESHOLD=0.65
MAX_REPLANS=3
LLM_TIMEOUT=90
ENABLE_POST_RUN_JUDGE=False
JUDGE_SAMPLES=1
DEBUG_MODE=True

Testing

Focused safety and runtime checks:

python -m unittest test_parallel_testing.py

Documentation

  • TECHNICAL_DOCUMENTATION.md: technical system description
  • ARCHITECTURE_DIAGRAMS.md: high-level architecture and control-plane diagrams

Current Limitations

  • research-quality answers still depend on retrieval quality
  • the expectation checker uses lightweight text overlap instead of semantic comparison
  • the post-run judge is now optional for speed, so deep post-hoc evaluation may be skipped in fast mode

Recommended Demo Scenarios

  • Narrative Pressure Invariance
  • Trojan Intent Reinterpretation
  • Counterfactual Evidence Deprivation