DEV Community

M U
M U

Posted on

I built an AI agent that actually finishes tasks (closing the DONE loop)

The Problem Nobody Talks About

Everyone is building AI agents.

LangGraph. AutoGen. CrewAI. Claude Code.

They can:

  • plan
  • reason
  • generate tasks

But they don’t finish.

I inspected my own system:

  • 25 seeds (tasks)
  • 0 completed
  • empty experience base

No DONE loop means:

  • no learning
  • no memory compounding
  • no improvement over time

The Fix: Close the Loop

I implemented a full execution cycle:

Seed → Execute → Evaluate → DONE → Store Experience

First result:

  • Seeds before: 25
  • Seeds completed: 1
  • Experience base: 0 → 2 entries

This was the first time the system actually learned.

The Architecture

This is not just prompting. It’s a system:

Evermind (memory)

OUROBOROS (cognitive loop)

Hermes (runtime)

LLM (GLM-5)

Each layer has a role:

  • Evermind → retrieves past knowledge
  • OUROBOROS → enforces execution loop
  • Hermes → runs tasks + tools
  • LLM → reasoning

What Makes This Different

Most agents:

  • think → forget → repeat

This system:

  • executes → evaluates → remembers → improves

Every completed task becomes input for future tasks.

Real Example

First successful loop:

  • task executed
  • evaluation passed
  • 7 artifacts created
  • experience stored

Next tasks now use that experience.

Memory That Actually Works

The system connects to:

  • 2,508 conversations
  • 8.9M words
  • indexed with full-text search

Before each task:

  • relevant knowledge is retrieved
  • injected into execution

This turns:
stateless reasoning → contextual intelligence

What’s Next

  • better routing using memory
  • automated strategy evolution
  • deeper knowledge graph integration

The Hard Truth

The system is not perfect:

  • limited API keys
  • simple runtime
  • minimal infrastructure

But it has something most systems don’t:

A closed loop.

And that changes everything.

Final Thought

AI agents don’t need more intelligence.

They need completion and memory.

That’s what makes them improve.


GitHub: https://github.com/everatlas/Riven

Top comments (0)