Amplifier Stories

The System That Builds Itself

Amplifier's adversarial bundle factory

BundleWizard · v0.4.0 · Active

The Frame

A bundle is a composition, not a monolith — and hand-building one is where the bugs hide

An Amplifier bundle wires behaviors, agents, context, and tools together with clear boundaries. It is not one large document that tries to do everything.

So who should assemble these pieces — and can it be trusted to?

“A bundle is a composition of capabilities — behaviors, agents, context, tools — wired together with clear boundaries.”

philosophy.md, line 7

The Proof

BundleWizard is real and it works — and one author built the whole thing

A bundle generation and improvement factory for the Amplifier ecosystem, built and verified in the open over eleven days.

A working factory. But what stops a factory from mass-producing bad output?

145

commits, one author (M. Jabbour)

4

merged PRs (#1–#4)

197

passing tests (175 + 22 bridge)

11

days: Mar 16–27, 2026

The Tension

Generators hallucinate plausible artifacts — a factory with no gate ships broken bundles

They produce plausible-looking artifacts with subtle bugs. And the danger hides in plain sight, exactly where nobody looks.

So the factory needs a mechanism that assumes its own output is wrong.

“Generators hallucinate — they produce plausible-looking artifacts with subtle bugs.”

philosophy.md, line 54

“‘Simple’ bundles are where the bugs hide.”

convergence-criteria.md, line 3

The Mechanism — the loop

The execute stage is an adversarial loop that repeats until it converges

One cycle inside BundleWizard's six-stage flow (explore → spec → plan → execute → verify → finish). If the critique says “NEEDS CHANGES” the refiner runs; “APPROVED” skips straight to evaluate.

The loop only works if the critique is honest. So who reviews the work?

1

Generate — produce the artifact

2

Critique — adversarial review

3

Refine — only if changes needed

4

Evaluate — score against the gate

The Mechanism — fresh eyes

The critic runs with fresh eyes — none of the generator's context

Spawned with context_depth="none", the critic has zero context from the generation process. That is what makes the review genuinely adversarial instead of a rubber stamp.

Honest critique catches bugs — but when is a bundle actually done?

“You are the adversarial reviewer… you should have been spawned with context_depth="none", meaning you have NO context from the generation process.”

bundle-critic.md, line 43

The Mechanism — the gate

Nothing ships until it clears a hard three-level formula

“Converged” is an objective threshold, not an opinion. The loop runs up to 10 iterations (typically 2–5 passes), and the best result is always checkpointed.

Because the gate is machine-checked, the human no longer has to stand at it.

L1 — Structural

must PASS

L2 — Philosophical

score ≥ 0.85

L3 — Functional

score ≥ 0.80

converged = (L1 == PASS) AND (L2 ≥ 0.85) AND (L3 ≥ 0.80)

The Payoff

Amplifier spawns BundleWizard when it detects a missing capability

Amplifier performs the explore phase directly — it already knows what it needs — then launches the autonomous continuation. Human approval is gone, but the quality bar is not.

The machines hold the bar. That is how Amplifier builds its own extensions.

“Machine quality gates remain fully enforced — you skip human approval, not quality.”

README.md, line 66

The Takeaway

When machines hold the quality bar, autonomous means trusted — not unconstrained

Removing the human checkpoints does not remove the craft. The gates stay; the loop stays; the critic stays adversarial. What leaves is only the wait for a human to say yes.

Sources

Sources & Research Methodology

Data as of: July 21, 2026 · Feature status: Active (v0.4.0)

Repository: michaeljabbour/amplifier-bundle-bundlewizard (local clone /tmp/bundlewizard, HEAD e1c5f9f)

Research performed — every claim independently re-derived:

Commit history: git log --oneline | wc -l → 145 commits
Contributor: git log --format='%an <%ae>' | sort | uniq -c → 1 author (Michael J. Jabbour)
Date span: git log --format='%ad' --date=short | sort | uniq -c → 2026-03-16 to 2026-03-27
PR history: gh pr list --repo michaeljabbour/amplifier-bundle-bundlewizard --state all → PRs #1–#4 merged
Tests: python3 -m pytest tests/ -q → 175 passed; bridge/ with PYTHONPATH=src → 22 passed (197 total)
Composition: ls agents/*.md context/*.md recipes/*.yaml modes/*.md | wc -l; ls -d skills/*/ → 11 / 8 / 6 / 7 / 2
Mechanism & gates: sed -n / grep -rn over philosophy.md, convergence-criteria.md, bundle-critic.md, recipes/bundle-single-iteration.yaml, README.md, bundle.md

Gaps / framing notes: The phrase “even experts ship broken bundles” is narrative framing, not a repo quote; the “composition complexity / ecosystem blindness / no quality gates” triad is synthesized framing (each maps to real repo content). Commit count is the local HEAD; upstream may have advanced.

Primary contributor: Michael J. Jabbour — 145 / 145 commits (100%).