Home
A distributed action gateway that transforms, deduplicates, routes, and dispatches actions through a configurable pipeline of rules, providers, and state backends.
What is Acteon?¶
Acteon is a distributed action gateway built in Rust. It sits between your application and external services (email, SMS, Slack, webhooks, LLMs, and more), providing a powerful control plane that decides what happens to every action before it reaches the outside world.
The name draws from the Greek myth of Actaeon, a hunter transformed by Artemis into a stag — the very thing he pursued. Likewise, actions entering Acteon are transformed — deduplicated, rerouted, throttled, or dispatched — before they ever reach the outside world.
flowchart LR
A[Your Application] -->|Action| B(Acteon Gateway)
B --> C{Rule Engine}
C -->|Allow| D[Provider Execution]
C -->|Deduplicate| E[Return Cached]
C -->|Suppress| F[Block Action]
C -->|Throttle| G[Rate Limit]
C -->|Reroute| H[Different Provider]
D --> I[Email / Twilio / Slack / Teams / Discord / Webhook / SNS / Lambda / SQS / S3]Rule-Based Processing¶
Define rules in YAML to suppress, deduplicate, throttle, reroute, or modify actions. Rules are evaluated by priority and can match on any field of the action payload.
Event Lifecycle Management¶
Track events through configurable state machines with automatic timeout transitions. Group related events for consolidated notifications.
Pluggable Backends¶
Choose from Memory, Redis, PostgreSQL, or DynamoDB for state storage. Mix and match with PostgreSQL, ClickHouse, or Elasticsearch for audit trails. (Note: ClickHouse state backend was removed in v0.1.0; see Migration Guide).
Human-in-the-Loop Approvals¶
Require human approval before executing sensitive actions. Supports HMAC-signed approval/rejection URLs with configurable TTLs.
Multi-Step Task Chains¶
Orchestrate multi-step workflows where each step's output feeds into the next. Supports configurable failure policies and delay between steps.
LLM Guardrails¶
Use LLM-based evaluation to gate actions through AI-powered guardrails. Block or flag actions based on content analysis with configurable confidence thresholds.
Semantic Routing¶
Route actions by meaning, not just field values. Use vector embeddings and cosine similarity to match actions against topic descriptions in natural language.
AWS Providers¶
Native integrations for SNS, Lambda, EventBridge, SQS, S3, and SES with automatic STS credential refresh, cross-account role assumption, and LocalStack support for local development.
Enterprise Ready¶
Multi-tenant isolation, API key and JWT authentication, hot-reload for rules and auth config, graceful shutdown, and comprehensive audit trails.
Polyglot Clients¶
Official SDKs for Rust, Python, Node.js/TypeScript, Go, and Java. Every client provides dispatch, batch, rule management, and audit querying.
Simulation & Testing¶
End-to-end testing framework with mock providers, failure injection, multi-node scenarios, and performance benchmarks across all backend combinations.
Guides¶
In-depth guides that combine multiple Acteon features to solve real-world problems, from AI agent swarm coordination to production deployment patterns.
Agent Swarm Orchestrator¶
Acteon also ships acteon-swarm — a standalone multi-agent orchestrator that uses the Acteon gateway for safety enforcement but is a separate product surface. While Acteon core is an action gateway (rules, routing, audit), the swarm is an agent orchestration engine inspired by Karpathy's autoresearch pattern.
Build → Eval → Critique → Fix → Verify¶
Primary agents build your project, an eval harness scores it, adversarial agents critique across engines (Claude + Gemini), and recovery agents fix code — all gated by fitness scoring with automatic git revert on regression.
Cross-Engine Adversarial Review¶
Use a different AI engine for critique than for building — Claude primary with Gemini adversarial, or vice versa. Cross-model blind spots catch issues that a single model misses.
Autoresearch Primitives¶
Three Karpathy primitives: editable asset (workspace), scalar metric (eval score), time-boxed cycles. Plus a fourth: SWE-bench-style binary assertions auto-generated from adversarial challenges.
Acteon as Safety Layer¶
Every agent tool call flows through Acteon for policy enforcement — dedup, throttle, approval gates, audit trail. The swarm is the orchestrator; Acteon is the guardrail.
Quick Example¶
Define a rule to deduplicate emails and suppress spam:
rules:
- name: block-spam
priority: 1
condition:
field: action.action_type
eq: "spam"
action:
type: suppress
- name: dedup-email
priority: 10
condition:
field: action.action_type
eq: "send_email"
action:
type: deduplicate
ttl_seconds: 300
Dispatch an action:
curl -X POST http://localhost:8080/v1/dispatch \
-H "Content-Type: application/json" \
-d '{
"namespace": "notifications",
"tenant": "tenant-1",
"provider": "email",
"action_type": "send_email",
"payload": {
"to": "user@example.com",
"subject": "Welcome!"
},
"dedup_key": "welcome-user@example.com"
}'
Architecture at a Glance¶
graph TB
subgraph Intake
API[HTTP API / Swagger UI]
CLI[CLI / Clients]
end
subgraph Control Plane
RE[Rule Engine]
SM[State Machines]
GRP[Event Grouping]
APR[Approval Manager]
CHN[Chain Orchestrator]
LLM[LLM Guardrails]
end
subgraph Execution
EX[Executor with Retries]
PROV[Provider Registry]
end
subgraph State
MEM[(Memory)]
RED[(Redis)]
PG[(PostgreSQL)]
DDB[(DynamoDB)]
end
subgraph Audit
AUD_PG[(PostgreSQL)]
AUD_CH[(ClickHouse)]
AUD_ES[(Elasticsearch)]
end
API --> RE
CLI --> RE
RE --> SM
RE --> GRP
RE --> APR
RE --> CHN
RE --> LLM
RE --> EX
EX --> PROV
PROV --> Email[Email SMTP/SES]
PROV --> Slack[Slack]
PROV --> Twilio[Twilio SMS]
PROV --> Teams[Teams]
PROV --> Discord[Discord]
PROV --> WH[Webhooks]
PROV --> AWS[AWS SNS/Lambda/SQS/S3]
RE -.-> MEM & RED & PG & DDB
EX -.-> AUD_PG & AUD_CH & AUD_ESLicense¶
Copyright 2026 Penserai Inc. Licensed under the Apache License 2.0.