Self-Awareness
From the Outside In

A self-improving orchestrator built from standard parts

The Bet

Self-improvement isn't a model problem.
It's an architecture problem.

A self-improving AI seems to demand a new model. The self-driving orchestrator (ramparte/amplifier-bundle-self-driving) manufactures the equivalent of self-awareness from external structure alone — no new model, no bespoke framework.

So where does the "self" come from? Start with what it's actually made of.

Everything is bundles, hooks, and tools — nothing new.

The whole orchestrator is under 3,800 tracked lines, built in a single ~5-hour day. DESIGN.md 1.1: "Not a new LLM or fine-tune… Not an Amplifier kernel modification."

If it isn't the model doing the work, what is?

3,779
Total tracked lines (md/yaml/py/sh/json)
~5 hrs
8 commits, single day (2026-02-10)
0
New models or kernel modifications
The Real Problem

A model can't reliably tell if it's alive, stuck, or doing well.

Introspection isn't a dependable source of self-awareness. DESIGN.md frames metacognition "built from the outside in: not by making the model introspective, but by building external structures — state files, review loops, goal registries — that create the functional equivalent of self-awareness."

If self-awareness can't come from inside the model, it has to be built outside it. Here's the first piece.

Three zero-dependency modules give the system a survival sense.

705 LOC of pure amplifier_core primitives — every pyproject.toml declares dependencies = []. Token tracking exists because Amplifier has no built-in cumulative token count.

Survival is one axis. Quality is another — and it can't be self-graded.

1
tool-session-state — 364 LOC5 file-based tools: read/write state, append log, list/read tasks
2
hooks-heartbeat — 112 LOCWrites last_heartbeat on tool:post & content_block:end
3
hooks-token-tracker — 229 LOCAccumulates per turn; injects counts via status-context pattern

An adversarial reviewer never sees the worker's reasoning.

Both worker and reviewer run with context_depth: none. The reviewer sees only the task spec, the claimed output, and files on disk — and is read-only, so the quality signal is honest, not self-flattery.

State, heartbeat, review — all outside the model. What happens when the process dies?

The state files themselves ARE the recovery mechanism.

All state is file-based JSON in self-driving-state/ — no database. An ultra-thin 220-line PID watchdog watches liveness and restarts on death, up to 3 times.

Every layer lives outside the model, in inspectable structure. Now the whole design snaps into one shape.

# scripts/watchdog.sh MAX_RESTARTS=3 CHECK_INTERVAL=30 # seconds HEARTBEAT_STALE=600 # seconds # dead PID -> restart (max 3), # then status = "failed" # state: config.json, goals.json, # orchestrator.json, log.jsonl, # tasks/task-NNN.json

Five layers, five questions — each answered by structure.

The Maslow-style hierarchy (L1 Survival → L5 Growth) is enforced bottom-up before any task. Each layer answers a self-awareness question with external structure — not a smarter model.

None of these needed a smarter model. Only structure you can build today.

What To Keep

Build the equivalent of self-awareness today — from the outside in.

L1–L4 are built and integration-tested (Phases 1–4). L5 "Growth" — the wisdom store and self-proposed goals — is designed but explicitly aspirational: DESIGN.md marks it Future ("This is the hard one"). The honest frontier, stated as such.

L1–L4 built • L5 designed / future

Sources

Sources & Research Methodology

Data as of: repo state at last commit 2026-02-10 (ec31a18)

Feature status: Design + Phases 1–4 built and integration-tested; DESIGN.md header still reads "Design phase" (stale metadata). L5 Growth designed but aspirational.

Repository: ramparte/amplifier-bundle-self-driving (includes microsoft/amplifier-foundation@main)

Research performed:

Gaps: The design cites amplifier-distro "overnight build" numbers as prior art (NOT independently verified). No end-to-end overnight session was run — runtime performance claims are unverified. L5 Growth is designed, not yet built.

Primary contributor: Sam Schillace (sole git author, 8/8 commits; DESIGN.md "Sam Schillace + Amplifier design session"). Commits carry a Co-Authored-By: Amplifier trailer.

More Amplifier Stories