cognisagent

Zero-boilerplate agents with pluggable middleware and backends.

cognisagent is the application layer of the Cognis framework. It wraps cognis and cognisgraph into a batteries-included agent factory — configure once, get a production-ready agent with middleware hooks, storage backends, planning, and multi-agent collaboration.

Quick Start

[dependencies]
cognisagent = "0.1"
tokio = { version = "1", features = ["full"] }
serde_json = "1"

use cognisagent::config::DeepAgentConfig;
use cognisagent::create_deep_agent;

let config = DeepAgentConfig::default();
let graph = create_deep_agent(config)?;
let result = graph.invoke(serde_json::json!({"messages": []})).await?;

Middleware Pipeline

The Middleware trait provides four hooks around every model and tool call:

before_model  ->  LLM call  ->  after_model
before_tool   ->  Tool exec  ->  after_tool

Built-in middleware:

Middleware	What it does
`FilesystemMiddleware`	File read, write, list, glob, grep
`MemoryMiddleware`	Inject persistent memory into context
`SubAgentMiddleware`	Delegate tasks to isolated sub-agents
`SummarizationMiddleware`	Auto-summarize when context window fills
`PlanningMiddleware`	Plan-then-execute with step tracking
`ContextMiddleware`	Dynamic context injection
`RateLimiterMiddleware`	Token bucket rate limiting
`LoggingMiddleware`	Structured logging with PII redaction
`PatchToolCallsMiddleware`	Auto-correct malformed tool calls
`SkillsMiddleware`	Load and invoke custom skills

Storage Backends

Backend	Description
`StateBackend`	In-memory (default, zero setup)
`FilesystemBackend`	Persist sessions as JSON on disk
`SandboxBackend`	Isolated execution with resource limits

More Features

Tool Registry — Centralized tool management with permission levels, call counting, enable/disable, and filtering.

Planning — Plan-then-execute pattern with step-by-step tracking and replanning.

Multi-Agent — Orchestrate multiple agents with shared state and message passing.

Presets — Ready-made agent configurations via a preset registry for common agent types.

Evaluation — Built-in evaluation framework for testing agent quality.

Health Monitoring — Diagnostics, profiling, and alerting for production deployments.

Part of the Cognis Workspace