Runtime Architecture

Deterministic
Agentic Execution

FemtoClaw’s runtime mediates between generative reasoning systems and deterministic physical or virtual infrastructure. It enforces capability policies, validates every tool call, and grants your agents true sovereignty without sacrificing safety or performance.

// Create an industrial agent
use femtoclaw_runtime::prelude::*;
fn main() -> Result<(), Error> {
let agent = Agent::new(Config::industrial());
agent.enforce_policy(Policy::Strict);
agent.run_managed()
}
Core Principles

Why FemtoClaw?

Deterministic Control

Every agent action is deterministically executed by the runtime, ensuring reproducibility and correctness. FemtoClaw enforces deterministic execution control and strictly manages side effects【93068253089653†L219-L225】.

Protocol Validation

The runtime mediates between inference and system authority via a protocol‑validated interface, validating each tool call’s structure and parameters【93068253089653†L219-L225】.

Capability Gates

Agents do not interact with the system directly. All operations are passed through capability gates that enforce deny‑by‑default policies and memory safe execution【93068253089653†L219-L225】.

Memory Safety

Implemented in Rust, FemtoClaw prevents undefined behaviour, protects against buffer overflows and data races, and provides memory‑safe runtime guarantees【93068253089653†L219-L225】.

Production Performance

With sub‑millisecond startup time and negligible overhead, FemtoClaw meets the demands of enterprise applications and industrial deployments【93068253089653†L334-L340】.

Extensible Architecture

Extend the runtime via pluggable capability modules, custom inference providers, memory backends and protocol extensions to meet your organisation’s unique requirements【93068253089653†L366-L373】.

Execution Model

Inference → Runtime → System

Probabilistic Inference

Generative models produce high‑level plans and candidate actions. These cannot interact with the environment directly; they must pass through the FemtoClaw runtime for validation and gating【93068253089653†L233-L241】.

Runtime Authority

FemtoClaw receives the candidate actions, validates their structure, applies capability policies and ensures they are deterministic. Only then are they forwarded to the system【93068253089653†L233-L241】.

Deterministic System

Once validated and authorised, the tool call executes on your underlying infrastructure. Observability and error reporting feed back into the runtime, never into the inference system【93068253089653†L233-L241】.

Runtime Subsystems

Building Blocks

Agent Core

The core of every agent, responsible for lifecycle management, event scheduling and orchestrating interactions with other subsystems【93068253089653†L245-L269】.

Brain Interface

Connects the agent to generative inference providers (LLMs or other reasoning engines) via a strict protocol, shielding your infrastructure from raw model output【93068253089653†L245-L269】.

Capability Gate

Implements a deny‑by‑default policy for tool execution and ensures every system call is explicitly authorised【93068253089653†L245-L269】.

Protocol Validator

Strictly validates JSON tool call structure, ensuring that no natural language ever reaches the execution layer【93068253089653†L245-L269】.

Memory Subsystem

Provides WAL durability, crash recovery and deterministic state replay for long‑running agents【93068253089653†L245-L269】.

Observability Layer

Emits structured logs and metrics for each decision and system call, ensuring you can audit and trace every step【93068253089653†L245-L269】.

Quick Start

Get Running

Install FemtoClaw via Cargo and run your first agent using the built‑in CLI. The runtime is distributed as a single binary without external dependencies【93068253089653†L381-L389】.

// Install and run
cargo install femtoclaw
femtoclaw run --brain local
femtoclaw run --brain remote --policy Strict
Performance & Deployment

Engineered for Production

  • Startup: less than 3 ms and memory footprint: under 5 MiB【93068253089653†L334-L340】.
  • Supports Linux, macOS, Windows, containers, edge devices and fully air‑gapped systems【93068253089653†L381-L389】.
  • Zero dependency, single binary deployment – bring your own inference models or use remote providers【93068253089653†L381-L389】.
  • Compliance suite verifies determinism, performance and safety against a comprehensive test suite【93068253089653†L352-L360】.
Further Reading

Dive Deeper into the Runtime

Explore an in‑depth breakdown of each core subsystem powering FemtoClaw’s deterministic execution authority. Understand how the agent core, brain interface, capability gate, protocol validator, memory subsystem and observability layer work together to enforce agentic sovereignty【93068253089653†L245-L269】.

Runtime Subsystems arrow_outward