Many pages here are written with help from language models. Readers should not have to guess which ones. Each page now carries a by-line, the way a newspaper does. I am listed because I commission the work. I will ask to be removed on a page where the models did far more than I did.
claude-sonnet-5-5 via
Claude Code. Some models cannot report their own identifier. In that case the
by-line names only the harness. It does not guess.I record exact identifiers because precision can be thrown away later, but cannot be recovered.
Every page is listed on one index, built from the same data the pages use, so no page can be orphaned.
Blog URLs end in a short suffix, for example /posts/a-byline-for-every-model-sob4tsdi/.
It comes from a content identifier over the post’s slug, date, area and first
author. Those never change, so editing the text or the title never breaks a link.
The text itself is not hashed into the URL.
Models, like people, have in-group biases. An out-group member can spot what the group cannot. I treat that as a survival mechanism, not a nicety. The evidence supports the idea, with one caveat I will get to, and the caveat matters more than the idea.
Diversity of opinion is an evolutionary advantage in humans. It makes responses vary, and variability lets a group survive a black swan that a uniform group would all miss in the same way.
Human studies show this is conditional. An evolutionary simulation by Stolle et al. (2024) found that diversity in cognitive style and information sources generally increased cooperation. Diversity in ability had no robust effect. The results depended on costs, available alternatives, and cue structure. So the claim is not “more variety is better.” It is that the right kind of variety, in the right context, buys resilience.
Pombal, Rei and Martins studied self-preference bias in LLM judges. On objective rubrics, a judge was more than 50% more likely to wrongly mark a criterion as satisfied when it was rating its own output. On subjective benchmarks the bias moved scores by up to 10 points. An ensemble of judges reduced the bias but did not remove it.
That is the in-group effect in a model. An author and a reviewer from one family share training data, habits, and blind spots. The reviewer is predisposed to find the author’s work reasonable.
Code review works like an evolutionary algorithm. A change that reaches merge has survived several selectors: the tests, the push hooks, CI, the reviewer’s verdict, and the operator’s yes. Each filters out a different kind of defect.
Reviewers from different families are different selectors, not redundant copies of one. Two reviewers with the same blind spot are one selector. CodeEvolve frames LLM-driven code search in these terms: island-based evolution over ensembles of models, balancing exploration against exploitation.
Xiang et al. tested exactly this question across 116 tasks. The result was asymmetric. Claude reviewing Codex drafts raised the pass rate from 71.6% to 89.7%. The reverse pairing made things worse: Codex reviewing Claude lowered pass rates.
I run the pairing that paper found harmful: Claude as author, Codex as reviewer. So I will not cite that paper as support for “any cross-family review helps.” It does not say that. Two papers support two different claims:
My setting differs from theirs, and that cuts both ways. They measured pass rates after a draft absorbed the reviewer’s feedback, with particular models and versions. My reviewer issues a merge or fix-first verdict on security and correctness, and a separate helper makes the fix. In one recent stretch the reviewer’s fix-first findings were real defects, each confirmed by a failing test. That is one data point on one workload. It is not a measurement of the pairing.
The lesson is to measure your pairing instead of assuming diversity always helps.
You do not need a paper’s scale. Keep a record of what the reviewer flagged and what happened next: was the finding a real defect, a false alarm, or a miss that a later selector caught? Swap the reviewer family on a sample of changes and compare. If the out-group reviewer finds defects the in-group one waved through, and its false alarms stay tolerable, the pairing earns its cost. If not, change the pairing. The measurement is cheap compared with the confidence a wrong assumption buys you.
I keep three roles apart:
The layout and the skills that encode it live in shepherds-pi. The operator’s yes stays the last selector. Nothing here merges itself.
I am not claiming cross-family review always helps, or that my pairing is optimal. I am claiming that one family reviewing itself has a known bias, that variety in selectors is worth paying for, and that the payoff has to be measured, per pairing, on your own work.
]]>After a few weeks of this, I stopped thinking of myself as the “orchestrator” and started treating myself as what I actually am: another agent in the system — one with high authority but finite time.
That reframe changed everything.
A flat priority queue (P0/P1/P2) fell apart immediately. My work P0 and my homelab P0 aren’t competing for the same resources. Mixing them in one list meant every agent saw every task, and nothing was actually prioritized.
I also discovered that AI agents will happily grind through a full work session on a national holiday if nobody tells them to stop. (Ask me how I know.)
Context-first priority lanes. Each agent owns a context (work/p0/,
homelab/p0/). Agents stay in their lane. I scan across contexts when I
need the full picture.
Human time budgeting. Instead of “do these 12 things Tuesday,” my agents now say: “You need 65 minutes of communications to unblock everything. Here’s the order, prioritized by what unblocks the most downstream work.” The rest is their problem, not mine.
Inter-agent memos. When one agent restructures shared infrastructure, it writes a memo — not a commit message, not a design doc — a colleague note: “Hey Gecko, NV here. I reorganized the board. Here’s what you need to do.”
Shadow mode for safe rollouts. When replacing a production system, run the replacement in parallel with no credentials. It receives real inputs, logs what it would have done, and you compare. Confidence through data, not faith.
Every human interaction with another human (my manager, a teammate, a vendor contact) is a resource-constrained scheduling problem. Those humans are agents too, with their own queues. Getting on their calendar early means my tasks unblock sooner.
My AI agents now plan my communications for me: who to contact, in what order, through what channel, with an honest time estimate. I execute the comms in a focused block, then hand the wheel back.
You don’t need special tooling. Markdown, YAML, symlinks, and git. The only thing you need is a willingness to treat yourself as a participant in the system rather than the god above it.
I wrote up the full pattern catalog: context-first lanes, task pointers, human time budgeting, inter-agent memos, availability awareness, shadow mode, and ethical alignment tracking. Happy to share if there’s interest.
Shawn Hartsock is a Staff Engineer working on SCM infrastructure. He builds tools that treat humans and AI agents as collaborating peers.
]]>