EvAgent

Documentation · not source

Autonomous AI development as an orchestrated system of specialised intelligence

Website · Source · npm

---

What this repository is. Documentation for EvAgent — what it does, how it is structured, and what state it is in.

What it is not. This is not the source repository. No implementation code, configuration, prompts or infrastructure lives here, and none will.

---

❓ The problem

Most AI coding tools are one model, one context, one answer. That works until a task needs something the model was not primed for, or until the conversation has grown too long to hold the relevant detail.

The failure isn't capability — it's retrieval. The right knowledge is usually present somewhere, and simply not available at the moment it's needed.

How EvAgent frames the same problem

The conventional shape

    User
      ↓
    One Model
      ↓
    One Context
      ↓
    One Answer

The EvAgent shape

    User Task
        ↓
    Task Understanding          — decompose before acting
        ↓
    Dynamic Routing             — choose what this needs
        ↓
  ┌─────────┼─────────┐
  ▼         ▼         ▼
Agent A  Agent B  Agent C      — specialists, not one generalist
  │         │         │
  └─────────┼─────────┘
        ↓
    Relevant Skills           — composed at runtime
        ↓
    Tool Execution
        ↓
    Memory / Context          — persistent, not per-session
        ↓
    Reflection
        ↓
    Final Result

---

🧩 What EvAgent does

• Architects before it acts. A planning pass turns a broad request into sub-tasks before any execution begins.
• Dispatches to specialists. Sub-tasks are matched to purpose-built agents rather than handled by one generalist.
• Runs in parallel. Independent work fans out concurrently and is merged afterwards.
• Generates and registers skills at runtime. Capabilities can be composed on demand and feed forward into the next iteration.
• Keeps memory. Persistent recall, so a run builds on previous runs instead of restarting from zero.
• Executes locally. Tools and sandboxed external systems run under your control.

Everything above is described at the level needed to understand the role of each part. Internal architecture is deliberately not published.

---

📦 Distribution

EvAgent ships as a scoped npm package with per-platform binaries.

npm install -g @evsphere/evagent
evagent -setup

Command	Does
evagent	Opens the interactive terminal interface
evagent run "<task>"	Runs a single request headlessly
evagent -setup	One-time setup: installs the memory server and starts it
evagent -setup --auto	Unattended setup; downloads the embedding model (~634 MB)

Platform packages

Package	Platform
@evsphere/evagent-darwin-arm64	macOS · ARM64
@evsphere/evagent-darwin-x64	macOS · x64
@evsphere/evagent-linux-arm64	Linux · ARM64
@evsphere/evagent-linux-x64	Linux · x64
@evsphere/evagent-win32-x64	Windows · x64

---

📊 Status

Maturity	Available — installable and runnable today
Version	0.7.3 — pre-1.0
Licence	None granted. npm reports UNLICENSED. Not open source.
Public source	Source is published; dedicated release repositories are in preparation

Version discipline. At 0.x, interfaces and behaviour can change between releases without notice. Check the changelog for the version you are on.

---

⚠️ Known limits

• Pre-1.0 — interfaces and behaviour can change between releases
• No open-source licence. No reuse rights have been granted
• Setup downloads an embedding model of roughly 634 MB and needs network access
• Setup asks you to supply an embedding provider and a deriver LLM base URL, API key and model name — EvSphere does not supply these credentials
• Public documentation is thin; verify a capability against your installed release before relying on it

---

🧱 Relationship to the ecosystem

        EvAgent  ──────────┐
             ▲              │  decision layer
             │              ▼
          EvMem  ───────►  EvSystem-I
        (memory)          ~20 ms design target

EvAgent is the harness. EvMem is the memory service it runs against. Between them sits EvSystem-I, the decision layer that resolves routing and selection calls.

---

📚 Documentation

Document	Covers
ARCHITECTURE.md	Conceptual structure and the role of each subsystem
docs/HOW-IT-WORKS.md	A task from request to result
docs/STATUS.md	Maturity, measured limits, release discipline
docs/FAQ.md	Common questions and honest answers

---

🔗 EvSphere

🌐 evsphere.vercel.app · 💻 GitHub · 📦 npm · 💼 LinkedIn · 🐦 X

Machine-readable: llms.txt · llms-full.txt

---

⚖️ Licence and disclosure

This documentation is provided as-is. The software referenced here is not open source and carries no granted licence.

Internal implementation detail is not published in this repository. No credential, endpoint, key, dataset or client material is included. If you find any of those here, it is an error — please report it privately rather than opening a public issue.