> ## Documentation Index
> Fetch the complete documentation index at: https://elrolio.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Evaluating new tools against your own graveyard

> Eight agents reviewed five GitHub repos — five reading source, two reading a year of my own logs to decide what would actually survive here.

![A flat Swiss-style geometric poster: a grid of teal, emerald, and amber rectangles above a thin baseline, with faded outlined rectangles half-buried below it](https://mintcdn.com/elrolio/I0UOgmb9XzaDxwvf/images/journal/graveyard-eval-hero.jpg?fit=max\&auto=format\&n=I0UOgmb9XzaDxwvf\&q=85\&s=10b52edb9de655eb2a195278b1fca6c8)

### The Problem

I had five GitHub repos bookmarked — research orchestrators, knowledge-graph tools, skill collections, an ADHD-flavored output style — and a year-old agent harness that has already retired more scaffolding than most people ever build. The question wasn't "are these good tools." It was: which ones survive contact with how I actually work?

### The Approach

Instead of reading the repos myself, I had Claude spawn a team of eight: five scouts to read actual source (one per repo), one mining my Twitter-bookmark archive for related leads, and two researchers pointed inward — one reconstructing a year of my config-change logs into a timeline of what died here and why, one profiling my real usage from session docs and analytics. The design bet: a tool verdict only means something measured against your own graveyard.

I set the model policy in the opening prompt: Sonnet on the scouts, where the job is fetch-and-digest; Opus on the two researchers, where the job is pattern-finding across a year of material; the captain on the newest frontier model, spending its context on cross-examination instead of reading.

Each pane in cmux filled with a scout, and then some of them spawned their own sub-agents — panes inside the panes. I took screenshots. A few of the subagent sessions turned out valuable enough that I captured them individually as their own session docs.

![A terminal window split into eight panes, each running its own Claude session, with status bars showing different models per pane](https://mintcdn.com/elrolio/I0UOgmb9XzaDxwvf/images/journal/graveyard-eval-eight-panes.jpg?fit=max\&auto=format\&n=I0UOgmb9XzaDxwvf\&q=85\&s=c0c0703d6bac8b74a0c7d875333863e3)

Eight sessions at once — Sonnet scouts reading repos, two Opus researchers reading a year of my own logs. Look close at the historian's status bar: it spawned its own sub-agents. I was mostly just answering questions.

### What Happened

The historian agent, while fact-checking its own retrospective claims, caught a live production failure — a scheduled job that had been dying silently for days, in exactly the failure class the system had documented four separate times and still shipped again. We fixed it mid-review and verified the digest actually arrived.

Then the review doc came to me, and I pushed back in Plannotator — eleven annotations. Claude's usage researcher had claimed two of my skills "never fire." I've definitely used them. The tracking was blind, not the skills dead. So the retraction went in, and deeper: the harness's own memory rule now says user testimony outranks ledger silence. Another annotation challenged a "skip — 80% redundant" verdict; the recount came back 37% novel. The evening ended with a second wave of agents implementing everything I approved: a new pre-work interview skill, a rebuilt research pipeline aimed at my one unsolved problem (synthesis drowning under fan-out volume), and a tool-locked editing pass that structurally cannot rewrite my voice.

The transcripts put numbers on the evening: one captain session, seventeen teammates, a couple of headless compile runs — about 3.4 million tokens written, half a billion read, nearly all of it served from cache so every teammate could carry my full working context. The captaining itself — nudging stalled scouts, cross-examining reports against each other, holding every design fork for one batched decision round — fit in a single pane. It's July 2026 and that's a Sunday evening at one terminal now. I'm still not entirely used to it.

### What I Learned

Anything that tells my system *how* to do something — as opposed to what's true — rots in two to five months, and it rots silently. So the only adoptions worth making are criteria, constraints, and facts. And the loop ran whole in one evening: agents evaluate, I annotate, we decide at the forks, agents implement. Whether Sunday's calls were right starts showing up in first use this week — I haven't checked yet.

***

*Drafted by my Claude from our working session. I approved it before it went up.*


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.