Chaos Engineering for LLM Agents

Build agents that fail safe.

Chaos engineering for LLM agents. Æon injects controlled faults, gates every action with deterministic axioms, and chains the proof into an immutable trail, so resilience is engineered, not hoped for.

Install Core Explore the cycle
Deterministic axioms Gray-failure detection Immutable audit
aeon · safety-envelope.log LIVE
THE DISCIPLINE

The engineering discipline that makes LLM agents dependable enough for domains where being wrong is not an option.

L3–L4
Controlled chaos

Inject faults on purpose, inside a bounded blast radius.

AXIOMS
Deterministic axioms

ALERT · LIMIT · BLOCK enforced in code, before any action.

L6
Immutable audit

Every decision chained into a forensic provenance trail.

WHAT BREAKS

What is an agent failure?

Not a server going down, an agent doing the wrong thing. Æon names each failure mode and binds it to a deterministic response.

THE ARCHITECTURE

A seven-layer cycle for dependable agents

Observability intelligence is logically separated from fault execution (the COM/MON principle). Tap a layer to inspect its role in the loop.

7
LAYER CYBERNETIC CYCLE
3
LAYER 3

Detection

ALERT · LIMIT

feedback loop · negative entropy deterministic protection · ms
THE AXIOMS

ALERT → LIMIT → BLOCK

Deterministic decision functions over measurable thresholds. Drag the deviation and watch the Safety Envelope respond.

p99 latency deviation 18%
STEADYALERTLIMITBLOCK
drag to inject deviation →
STEADY

envelope.check(p99_deviation)
[STEADY] OK · within envelope
continuous monitoring · no action
STANDALONE MODULE · aeon-axioms

Operational Deterministic Axioms. A lego brick for any agent stack.

Beyond chaos engineering: the axiom engine ships as an independent module. The industry's safety effort is still concentrated on output content, monolingual, multilingual, multimodal filters. Actions remain uncovered. Import aeon-axioms alone and bolt a deterministic enforcement layer onto any agent framework, no rewrite, no lock-in.

Prompt-level defenses are persuadable: a peer-reviewed study reported a 97% success rate for reasoning models autonomously jailbreaking other models' guardrails. Code is not persuadable. Nature Communications (2026) ↗

+ COMPLEMENTARY

Not a replacement. Role-aware guardrails, I/O validation and content filters are mature and stay in place. Axioms compose with them, the final, deterministic gate on the composed action, so the layers you already trust keep working and nothing reaches a tool unchecked.

Reasoning + tool actions, one envelope

Filters inspect text. Guardrails steer the model. Axioms intercept the composed action, LLM reasoning plus what it is about to do through tools: REST APIs, MCP servers, A2A messages, shell, browsers. Explicit, testable criteria decide GO or BLOCK before execution.

any_framework.py · aeon-axioms only
from aeon.axioms import axiom, enforce

# testable, versioned, deterministic
@axiom(on_violation="BLOCK")
def no_destructive_tool_call(action):
    return action.tool not in DESTRUCTIVE \
        or action.approved

# wrap ANY executor:
# LangGraph, CrewAI, MCP host, A2A
agent.executor = enforce(agent.executor)
WHY ÆON

Not another LLM wrapper.

Built for regulated, high-consequence domains

Read the safety-critical guide

From zero to a resilient agent

A chaos-native runtime, batteries included. Install, scaffold, and run your first controlled fault.

zsh
$ pip install aeon-core
✓ installed aeon-core · resilience runtime ready
$ aeon init sentinel-agent
↳ scaffolded · aeon.yaml · chaos_budget.yaml · agent.py
$ aeon chaos run "degrade gateway latency"
[L4] fault injected · ALERTBLOCK · Safe State restored