Confidently Wrong

Multi-agent investigation that ships proof

The Problem

One agent reading a codebase hands you a confident story

“Plausible narratives can be confidently wrong; the agent doesn't know what it didn't look at.”

Parallax Discovery is an Amplifier bundle that exists to solve exactly this: a single-agent read feels authoritative, yet it is incomplete and sometimes flatly wrong.

So the first question is whether many agents can actually go deeper —

Proof — The foundational investigation

Multi-agent investigation really can go deep on a complex system

Parallax's foundational study of the Amplifier ecosystem ran a full multi-agent investigation — and produced a mountain of concrete artifacts.

But depth alone turned out not to be the whole story —

18

agent investigations

3

waves

80+

artifact files

10+

DOT diagrams

The Complication

Depth alone isn't the point — no single view was complete

In that same run, each independent team surfaced findings the others missed. A big investigation is still blind in spots unless you deliberately look from multiple angles.

That blindness is what Wave 1 is built to attack —

Mechanism — Wave 1 Discovery

Triplicate teams attack every topic from three fresh angles

Wave 1 dispatches three independent agents per topic, each starting with fresh context (context_depth='none') so no shared narrative can bias them.

And when those three views collide, Parallax does something deliberate —

HOW

code-tracer

traces the mechanism in the source

WHAT

behavior-observer

observes what actually happens

WHERE

integration-mapper

maps where it connects

Mechanism — Disagreement as signal

When the agents disagree, that's data — not error

“Discrepancies are information, not failures; never reconcile by fiat.”

Disagreement is the signal that points investigators straight at what a single narrative would have quietly papered over.

But even the sharpest reading is still just reading —

The Turn — Wave 3

Then the antagonist stops reading and starts running code

A Wave-3 antagonist challenges the strongest claims with actual execution instead of trusting the page — the shift from provisional code-reading to execution-based verification.

And that shift is exactly what changed the answer —

The Payoff

Only execution turned a provisional read into confirmed proof

Running the code exposed a real deep_merge bug — list replacement silently dropping config — at two live production collision sites that 18 agents of code reading across two waves could not find.

The same execution pass also corrected a token estimate that reading got wrong.

2

live production collision sites the bug hit

18

code-reading agents that missed it across 2 waves

~23–26K

token estimate from code reading

~36–39K

corrected by execution

The Takeaway

The whole method ships as a reusable, governed Amplifier bundle

Parallax Discovery (v0.2.1) packages the pattern as a tested bundle — enforcing one hierarchy: execution > code tracing > documentation > assumption.

A repeatable method you can adopt — not a one-off war story.

53/53 tests pass

5 specialized agents (code-tracer, behavior-observer, integration-mapper, lead-investigator, antagonist)

4 wave modes (discovery, verification, adversarial, synthesis)

3 human approval gates between waves

Sources

Sources & Research Methodology

Primary source: bkrabach/amplifier-bundle-parallax-discovery — cloned from GitHub to /tmp/parallax-repo (origin git@github.com:bkrabach/amplifier-bundle-parallax-discovery.git). Not present locally under /home/ramparte/dev/ANext.

Feature status: Released and maintained; current version 0.2.1.

Research performed:

Version: git tag && grep -n version bundle.md && git log -1 — v0.2.1, commit 096884a
Commit history: git log --oneline | wc -l — 22 commits
Agents/modes: ls agents/ | wc -l (5), ls modes/ | wc -l (4)
Approval gates: grep -niE 'approval|required: true' recipes/parallax-discovery.yaml — 3 gates
Triplicate teams: grep -niE 'triplicate|agent-1|agent-2|agent-3' recipes/parallax-wave.yaml
Origin investigation (18 agents / 3 waves / 80+ artifacts / 10+ DOT diagrams): sed -n '1,20p' docs/plans/2026-03-07-parallax-discovery-design.md; README.md:183; context/methodology.md:104
deep_merge bug & two collision sites: grep -n 'collision sites' README.md context/methodology.md agents/antagonist.md
Token correction ~23–26K → ~36–39K: grep -n '23-26K|36-39K' context/methodology.md agents/antagonist.md
Tests: python3 -m pytest tests/ -q — 53 passed
Evidence hierarchy: sed -n '36,46p' context/methodology.md
Contributors: git shortlog -sne --all; PRs: gh pr list --state all

Primary contributor: Brian Krabach (author/maintainer, ~21 of 22 commits). External contributor: Samuel Lee (PR #1, bundle include-nesting fix).

Gaps: The origin-investigation findings are the repo's own self-reported results (cited from methodology.md, triplicate-pattern.md, design doc); the underlying investigation artifacts are not in this repo and were not independently re-verified. Internal wording varies on the token correction (“50% underestimate” vs “40% wrong”); the ~23–26K → ~36–39K figures are quoted directly.