Blog

The Tumult blog series covers the platform end-to-end, from first principles to advanced use cases.

Post Topic
Introducing Tumult What Tumult is, why it was built, and how it differs from existing tools
The AI Advantage How TOON’s token efficiency enables AI-native chaos analysis
Built-In Observability OpenTelemetry spans and resilience.* attributes — always on, zero config
The Plugin System Script plugins, native plugins, discovery order, and writing your own
The Experiment Format Deep dive into TOON experiment structure, providers, and tolerances
The Analytics Pipeline DuckDB + Arrow + Parquet: SQL over your chaos history
Kubernetes Chaos tumult-kubernetes: pod delete, node drain, deployment scaling
Statistical Baselines IQR, percentile, mean/stddev — replacing magic numbers with evidence
Compliance as Code DORA, NIS2, PCI-DSS 4.0 — experiments as regulatory evidence
Chaos Under Load Combining tumult-network and tumult-loadtest for realistic fault testing
The Full Span Waterfall Real SigNoz traces from a live Tumult experiment — the observability proof
Load During Chaos k6 load testing concurrent with fault injection — proving disruption in numbers
GameDay Is Here Coordinated campaigns with resilience scoring — 4/4 PASS, Score 1.00, COMPLIANT
The Road Ahead What’s delivered (Phases 0-8), what’s next, the full series index
Bring Your Own Agent tumult recommend --agent: Claude Code / Codex enhance recommendations and propose validated experiments
Chaos Without Root tumult-net: a userspace TCP chaos proxy — latency, throttling, corruption, and connection kills with no root, no tc, no docker
Your Agent Is Now a First-Class Tumult Operator The MCP server grows to 24 tools with annotations, structured output schemas, tumult:// resources — and a run→ingest→recommend loop that closes over MCP
ChaosGraph: Your Agent Stops Re-Reading Journals A typed knowledge graph over chaos data, built from journals on ingest and served to agents over MCP — a targeted answer stays bounded while journal-reading grows every run (~8× more compact per run, ~20× on store-wide queries)
Agentic Trajectories: Chaos Engineering for Agents That Think in Steps Multi-turn agent-graph fault modeling — inject a fault at one step and watch it cascade across a trajectory, with whole-trajectory contracts and four agentic subscores. The failure modes single-call testing can’t see
Windows Chaos, For Real: Native Faults Nobody Else Ships tumult-windows: native process-kill, CPU-stress, and firewall-blackhole faults — the fault domain no OSS competitor offers, validated live against a real Windows 11 guest
Where Compliance Breaks: Lineage on the Service Map declared service topology + compliance lineage: see where controls break on the map, which fault caused it, and where the recommender says to inject next — three live proof runs from the demo stack
Autopilot: Chaos Engineering That Decides — and Shows Its Work policy-gated autonomous fault injection: deterministic recommender + 13-rule safety gate, earned autonomy, decisions as graph lineage, parquet evidence archive — proven by three consecutive 12-proof demo runs

Table of contents

Tumult is open source under the Apache-2.0 license.