THE AI LAYER Protocol 973

The ADL Flow — how AEGIS thinks

Every call works its way through the tiers — the cheapest, fastest brain first, the cloud only for the genuinely hard cases. Most of the thinking stays local and free on repurposed hardware. Here is a visual on the All-Drama-Llama (ADL) flow — the models, the agents, and the custom-built engines they pull from.

Brain · 8 GB GPU Auditor · 4 GB GPU

This page is one architecture, shown six ways — the same tier ladder that carries a call from a Small Language Model up to a Large Language Model runs through every tab below, each one built to teach a different part of it.

Curious what this actually costs? The Flow tab has a watts-research interest list — we're measuring real per-query power draw, SLM vs LLM, over 3/6/12 months. Want to run any of this yourself? Request clone/pull access to the repo →

The Flow
Open-Source SLM
Local Inference
AEGIS Agents
Digital CNA
Isolation
tiers

Navigating the tiers

Navigating the tiers, step by step

how one event finds the right tier to answer it
  1. Each hexagon is a brain — full of events and the tier it operates in.
  2. The first stop is AEGIS itself and its stack: every tool answers its own task with its own capabilities and engines.
  3. When an agent's reasoning or a real decision is needed, the work starts navigating the tiers — beginning with the cheapest agent, ADL (llama3.2:3b), at Tier 0.
  4. If that agent needs support to make the call, it climbs to the next tier — Tier 1, DEEP LOCAL (llama3.1:8b) — firing up the agent responsible there.
  5. Each climb only happens when the tier below isn't sure, so the expensive tiers rarely run — cutting down on cost, compute uptime, and resources.
  6. In this example, the final tier is Tier 2: the Claude API (Haiku → Sonnet → Opus).
each press advances the event one tier
Ready. Pick a scenario and press Run the flow, or step through it yourself.
event
cost = watts × time
power0.15 kW
time0.00 s
rate$0.14/kWh
energy0.0000000 kWh
= cost$0.0000000
local tiers ≈ free · only watts
The formula is real (energy = power × time × rate). The wattage, per-query time and rate shown are representative estimates — illustrative, not live telemetry.
the math

The cost, on the board

The cost, worked out
equation for cost of watts per API call
Local — GeForce GTX 1070 · ≈150 W
energy = power × time
= 0.15 kW × (2s ÷ 3600)
= 0.0000833 kWh
cost = energy × rate
= 0.0000833 × $0.14/kWh
≈ $0.0000117
Cloud — one Claude API call
cost = (in × in-rate) + (out × out-rate)
= (1000 ÷ 1M × $3) + (600 ÷ 1M × $15)
= $0.003 + $0.009
≈ $0.012 / call
≈ 1000 in + 600 out tokens · Sonnet-class rates
One cloud call costs about 1,000× the same query run locally — and that gap compounds with every input and output token. Used correctly, small language models become a real answer to AI at massive scale. Can we use AI without disrupting the environment? An exponential 1,000× difference is a universal language.
Formulas are real. Figures illustrative: 150 W GPU · $0.14/kWh (US avg) · Claude ≈ $3 / $15 per million input/output tokens.
Watts research · want the real numbers? join the interest list

Interested in tracking the actual watt-cost of running various Ollama models on repurposed GPUs — real power draw, per-query and over time, model by model, so the figures above stop being estimates.

Email
Comments — optional
Which run interests you? — optional
fill in your email to send
shared

Agents, research & analytics

Riding alongside the tiers: the OpenClaw agent council, and two shared engines every model can draw on — Tavily for live web research, Little MAE for local behavioral analytics. Here is how they come together: a device probes 40 ports in seconds. Sentinel flags it and OpenClaw hands it to Bloodhound for recon. Bloodhound asks Little MAE — "is this host a regular or a stranger?" (a stranger, first seen three minutes ago) — and pulls Tavily for the latest on that scan signature. The model reads both as context, Scribe writes the finding in plain language, and the case is decided with local behavioral memory plus live web — most of it free and on-box, the cloud only if it is still unsure.

OpenClaw · agent council

multi-agent · each backed by a local model or the cloud
SentineltriageADL · local ✓ live
Bloodhoundrecondemo
Reverendmalware / REdemo
Scribereportingdemo
Overseerorchestratordemo

Sentinel is live on ADL today. The rest are one config line away from running on a local model too.

Tavily · web research

the live-web tool in the agentic stack · sponsored by Tavily

Tavily is a search-and-retrieval API built for AI agents — give it a query and it returns clean, ranked, source-cited web results. It's how AEGIS reaches current information the models were never trained on: fresh CVEs, IOC reputation, vendor advisories, and threat write-ups.

The pipeline: when a case escalates, the relay calls Tavily and the results are folded straight into the Claude prompt as grounding context — so the model reasons over live evidence instead of stale memory. It's a shared tool: DeepRoot escalation uses it today, and any tier or agent (Bloodhound being the natural one) can make the same relay call.

Little MAE · the Multi-Agent Engine in design

local analytics engine · device tenure · anomaly & spike scoring

Little MAE turns raw RF and endpoint data into readable graphical data sets — who's a regular vs. a stranger, off-hours anomalies, packet-rate spikes — and the models read those signals as context, the same way they read Tavily's web results. So the brains reason with local behavioral memory, not just the single event in front of them. Two engines, one prompt: Little MAE computes the numbers, the models interpret them.

hardware

One rig, two jobs

The big card — a GeForce GTX 1070 — is AEGIS's brain; the small card — an MSI single-fan GTX 1050 Ti — acts as the network auditor (Digital CNA), kept deliberately off the SOC path so a code or network audit never stalls live triage. Both are affordable, second-hand-friendly GPUs — and that's the point: anyone can afford to run capable local models, at a fraction of what the big cloud LLMs cost per query. Put it in perspective — when the hardware and the know-how are this accessible, AI stops being something only a few can wield, and we get a real say in how it's used.

8 GB GPU · the brain
local models
  • llama3.2:3b — ADL, resident always-on
  • llama3.1:8b — deep tier, loads on demand
4 GB GPU · the auditor
Digital CNA — off the AEGIS path
  • Digital CNA — Code & Network Auditor (qwen2.5-coder:3b)
open-source SLM

Two installs, five minutes

Everything ADL runs on is free, open-source, and this easy to put on your own machine. Here's the whole install for the two pieces — the model runner, and the agent layer on top of it.

What is Ollama?

the model runner

Ollama is an open-source tool that downloads, serves, and runs language models entirely on your own machine — no account, no API key, no data leaving the box. Pull a model once, and it answers every prompt after that straight from local weights on disk.

Windows PowerShell
Illustrative, not a live terminal — the commands are real, the output/timing is a demo.

Get Ollama

What is OpenClaw?

the agent layer

OpenClaw is the multi-agent bridge that sits on top of a model like Ollama — it gives an agent a soul (SOUL.md), a tool policy, and a gateway so it can hold a persistent presence instead of answering one prompt at a time. It's how AEGIS turns a local model into GLAUX, ERGANE, HESTIA, and Sandy.

Windows PowerShell
Illustrative, not a live terminal — the commands are real, the output/timing is a demo.

Get OpenClaw

What's an SLM?

summary

An SLM — Small Language Model — is a language model built small on purpose: a few billion parameters instead of hundreds of billions, sized to fit and run on a single consumer GPU instead of a datacenter. It answers less than a frontier model, but for a well-scoped job — triage, classification, a first pass on an alert — it's fast, free per query, and fully private, which is exactly why ADL leans on one as the always-on Tier 0 brain.

Run the same basic question through a cloud LLM instead and the gap is stark: a local query on a ~150 W consumer GPU costs a fraction of a cent, the same call to a frontier cloud model runs roughly 1,000× more — and that multiplier compounds across every token, every user, every day (worked out in the math above). Most of what people actually ask AI for — a quick lookup, a spelling check, drafting a short reply — is well inside what an SLM handles fine, locally. That's not just a cost story; it's a resource story: electricity, datacenter cooling water, and new datacenter buildout all scale with query volume. Point everyday questions at an SLM on hardware you already own instead of defaulting to the cloud, and that draw drops by roughly the same order of magnitude — for free, on your own machine, today.

local inference

How a local model actually answers

One machine, no internet in the loop. When you ask a local model a question, the whole round-trip happens on that box — the weights are already on disk.

Fox Echo · one machine · offlineno NIC in the loop
8 GB GPU
weights already on disk
idle
CPU hands the prompt to the GPU · the GPU runs the forward pass · the decision comes back — all on this box.
private · offline · free per query — every token is computed right here, 0 bytes leave the box

No internet involved

private · offline · free per query

Every reply is generated right there, from weights already on disk — nothing leaves the machine and no cloud is called. If a small model ever tells you it's "cloud-based," that's a hallucination: a model has no idea how it's actually deployed. That privacy is the whole reason triage runs locally first.

live chat · llama3.2:3b · offlineit can't see where it runs
running fully on this laptop — yet the model still insists it lives in the cloud

Giving it a soul

SOUL.md · the "soul paper" · CLAUDE.md

Because a model can't know what it is, you tell it — in its system prompt. The community calls this the soul paper: a SOUL.md that spells out an agent's identity, purpose, and guardrails, loaded at the very top of the prompt (used with local models, OpenClaw, and more). Claude Code's equivalent always-loaded context file is CLAUDE.md. That's exactly where you'd correct a small model that thinks it's "cloud-based" — you write what it is, what it's for, and where its limits are. It's where AEGIS gives each model its soul.

Literal definition (OpenClaw docs): “SOUL.md defines identity, tone, and boundaries.” The official template opens “You're not a chatbot. You're becoming someone.” — and it's the first file the agent reads before it acts. Official template: docs.openclaw.ai/reference/templates/SOUL
aegis agents

The defensive council — Brain · Orchestrator · Hands

The centerpiece: how AEGIS's own agents work. Same three roles — but here the hands are your trusted tools acting on your own network. They analyze and enrich; they never run untrusted code, so there's no throwaway VM.

alert → brain → OpenClaw → the council → trusted tools → a written verdict
Ready. Press Run the case to watch one alert move through the whole council.
incoming alert brain · local model OpenClaw · orchestrator council agents trusted tools
▸ the hands are your own trusted tools on your own network — they read and enrich, they never run untrusted code, so there's no throwaway VM here

Isolation by trust boundary

AEGIS's architecture isolates any agent or orchestrator that could come into contact with untrusted, external, or potentially harmful data — network captures, scanned URLs, third-party API responses, OSINT pipeline output — from the trusted core (the SOC / triage path). This isolation exists specifically to contain prompt injection: if an agent processing untrusted data is compromised via injected instructions, it cannot reach or influence the trusted orchestrator because it never had privileged access to begin with. This matches current OWASP LLM Top 10 guidance on least-privilege tool access and trust-boundary segregation between untrusted content and system instructions.

One alert, one loop

  1. A suspicious event hits the stream — Sysmon, RF, or a honeypot.
  2. The brain reads it and decides whether it's worth a closer look.
  3. The orchestrator hands it to the right specialist — Sentinel for triage, Bloodhound for recon.
  4. That specialist calls trusted tools (VirusTotal, nmap, Kismet, Tavily) and forms a verdict.
  5. The verdict lands on the case and the watchlist — your analyst reads a summary, not raw noise.
aegis · network auditor

Digital CNA

The Digital CNA — Code & Network Auditor — is part of AEGIS: the agent that audits your own network. It runs on the same three-role flow as the rest of the council — the brain reasons, OpenClaw orchestrates, and the work runs in a controlled environment.

Here's the reusable part: the same flow builds a CTF agent. Same architecture, different job. Below is the one diagram, applied twice.

CNA

The network auditor — part of AEGIS

CNA · your network

Auditing your own network

BRAIN
local model
OPENCLAW
orchestrator
AUDIT ENV
controlled
controlled environment · your network
~/agent · audit env

The same flow → a CTF agent

Swap the target and you've built something new. Point the exact same Brain → OpenClaw → environment loop at a challenge instead of your network and you have a CTF agent. One thing changes: a challenge runs untrusted code, so its environment becomes a disposable VM — isolated and destroyed after. Same flow, new job.

CTF

The CTF agent — same flow, opposite authorization

Same three-role loop, pointed two different ways. The architecture is machinery — technical and mechanical, and it doesn't know or care what it's pointed at, because it can't. That's exactly why architecture, infrastructure, isolation, a soul.md, and the flow itself matter so much when you build an agent or model system — they're what decide what the loop is allowed to touch. The loop itself has no say.

White hat · authorized
The brain
LOCAL MODEL
plans the next move
The orchestrator
OPENCLAW
decides & routes
The hands
DISPOSABLE VM
an authorized CTF challenge
Black hat · unauthorized
The brain
LOCAL MODEL
plans the next move
The orchestrator
OPENCLAW
decides & routes
The hands
TARGET HOST
no permission, no scope
IMPORTANT

Same code, same three-role loop. The only difference is authorization — a CTF box you were invited to break, versus a system you were never given permission to touch. That line is what makes one practice and the other a crime.

A VM tailored to the task

the orchestrator provisions the right toolbox before it starts

You tell the bot what kind of challenge it is; it spins up a fresh VM pre-loaded with only the tools that task needs — nothing more. The wrong tools never touch the box, and the whole thing is torn down when the flag is found.

web
burp · ffuf · sqlmap
pwn / binary
gdb · pwntools · ghidra
crypto
sage · python · hashcat
forensics
volatility · binwalk · wireshark
recon / osint
nmap · amass · theHarvester
reversing
radare2 · ghidra · angr

One challenge, one loop

  1. You ask the orchestrator to solve a challenge, and name its type.
  2. It provisions the tailored VM, then asks the brain: "here's the challenge — what's the next command?"
  3. The brain returns a plan; the orchestrator sends it over SSH into the disposable VM.
  4. The VM runs it and returns output; the orchestrator feeds that back to the brain, and repeats.
  5. The flag is written to the orchestrator's store — the VM is torn down, nothing persists in it.
isolation

Two machines, one wall

Isolation is a discipline we practice as we build — the security holy grail: keep the machine that runs untrusted code completely separate from the one you trust, so a compromise on one can never reach the other.

Machine 1 · trusted
Trusted
AEGIS HOST
SOC console
DeepRoot · ADL · Auspex
status only
isolation boundary
Machine 2 · isolated
Isolated
CTF AGENT
own host · user
own repo · own venv

Only status ever crosses that line — never code, never data. Everything below lives entirely on Machine 2.

Its own storage root

nothing shared with the AEGIS host
challenge_cache/
downloaded files
solves.db
flags · progress · notes
logs/
tool output · history
scratch/
wiped after each run
creds/ 🔒
encrypted platform keys
platform API
optional · outbound only

Execution sandbox untrusted code only

only spun up if a challenge binary must actually run
disposable container / VM per challenge — destroyed after each run
no network egress — except to the CTF platform itself
no shared mount — copy the file in, copy the result out
Blast radius = one VM. Even if a challenge binary is hostile, it can't reach your keys, your solves, or the AEGIS host — the sandbox is torn down and nothing it touched survives.

Coming next teaching series

We'll grow this into a lesson on both sides of the wall: how the boundary gets attacked — VM escape, egress abuse, shared-mount mistakes — and how we, as security students, learn to hold it. Isolation isn't a diagram; it's a practice you build the reflex for.