Agents System

background logic that actually does the dirty work for you. brainz ships with a built-in agent layer—watchdogs that sit between your prompts, memory, and training pipeline. they watch, score, rewrite, and fine-tune without you lifting a finger.

agents live here: backend/agents/

run them inline with inference or fire them off async in the background.

---

what they do

• hook directly into the llm lifecycle

• log every action (if you want it)

• fully scriptable + chainable

• can be swapped, extended, or rewritten in minutes

---

brain flow (high level)

new prompt → score it  
│  
├── good → push to vector memory  
└── trash → autotrainer fires → feedbackloop grabs session → promptoptimizer rewrites → fine-tune queued → memory updated → system evolves  

the longer it runs, the smarter it gets.

---

autotrainer

the muscle. auto-corrects the brain when it screws up.

• hooks into /api/llm/query

• scores output via utils/eval_metrics.py

• if score < threshold → logs prompt → triggers models/trainer.py live fine-tune → tags metadata for later digging

tweak it in core/config.py: DEFAULT_AUTO_TRAIN_THRESHOLD = 10

no cronjobs, no manual scripts. it literally teaches itself.

---

feedbackloop

the paranoia module. doesn’t trust a single answer, watches for failure patterns:

• long response delays

• repeated edits or retries

• same topic spammed over & over

if it smells “low confidence”: → embeds prompt → dumps it into vector memory → queues retraining → tags it for prompt rewrite

---

promptoptimizer

rewrites garbage prompts before they hit the model.

what it does:

• splits overstuffed queries

• cleans slang + weird abbreviations

• reshapes the question into llm-friendly structure

• keeps intent, boosts clarity

example: "wtf is mev?" → "explain mev (maximal extractable value) in ethereum"

under the hood: zero-shot scoring + regex hacks + semantic distance checks.

---

full live loop

user: "wtf is zk-snarks vs starks?" model: weak answer autotrainer: flags score < threshold feedbackloop: logs session promptoptimizer: rewrites to "compare zk-snarks and zk-starks in zero-knowledge cryptography" training: queued instantly memory: updated with new vector

→ you didn’t do anything. it just got smarter.

---

build your own agent

drop a file in backend/agents/:

class MyAgent:
    def run(self, prompt: str, response: str) -> dict:
        return {"action": "log", "priority": "medium"}

register it:

registry.register("my-agent", MyAgent())

agents can:

• write logs

• trigger webhooks

• mess with memory

• chain other agents

• queue training jobs

async, stackable, schedulable. your rules.

---

why it matters

agents = hands-free evolution.

• the more you use it → the better it gets

• no manual fine-tuning

• sloppy prompts cleaned up automatically

• output stabilizes over time

• the system runs itself, degen-style