Agents review agent code
Every change I push gets two fresh AI reviewers, never a fork of my own session, because a fork already agrees with me.
In the new world of AI agents,
I make sure agents get the right context at the right time, turning OK into amazing
I'm William “Billy” Bratz, a senior software engineer focused on context engineering: what production agents are handed when they answer, and what coding agents know when they write code.
−65%
Retrieval latency for a live agent, with 100% identical answers.
47% less
Cost per accepted PR when coding agents get domain context: $7.79 down to $4.12.
9 → 0
Blocker and major bugs in coding agent PRs, without vs with domain context.
0 → 6 of 8
Coding agent implementations done to our coding standards. Each rule pops up the moment the agent starts editing a file it applies to.
The problem
The model could have done it. It got the context too late, got too much of it, got the wrong piece, or never got the one fact that mattered. Each one has a better fix than a bigger model.
Too slow
Context arrives after the user stopped waiting
I predict the context an agent will need and load it ahead of time, with priority lanes so background loading can’t crowd out a live conversation.
−65% retrieval latency · identical answers
Too much
Bloat buries the signal and burns money
I find out what the model actually receives before cutting anything, so the savings come from real waste instead of guesses.
99%+ of a runaway request traced to one cause
Wrong
The agent sends the question to the wrong tool
The wrong tool means a confident, wrong answer. I make routing measurable, so any change to a tool, a prompt, or the model can be scored on whether every question still lands in the right place, before a customer finds out.
routing measurable across every tool
Missing
The one rule or past decision that would have changed the answer
A memory layer that files every rule where it’s needed and hands it to the agent the instant it touches the code that rule covers.
right rule read first · 95% of runs
Production agents
A customer-facing agent has a few seconds and a token budget per turn. I'm focused on the layer between the model and the systems it calls: how fast context arrives, how much of it the model carries, and how tools tell the model what they do.
Faster retrieval, same answers
65% faster at the median, with 100% identical output.
Context bloat, taken apart
A runaway token bill turned out to be over 99% inherited conversation history, already there before any tool ran.
Measure before you cut
Forwarding full tool schemas looked expensive. Measured: 0% more input tokens.
Tools the model can read
Output schemas on every tool, plus example questions that test whether an agent picks the right one.
Coding agents
Claude Code can program. It doesn't know why your system is shaped the way it is, which decisions are settled, or which mistakes already hurt you. I built The Vault to tell it, and proved the difference with blind tests on real tickets.
With the Vault vs without
+45%
accepted changes
Zero
blocker or major bugs
−47%
cost per accepted change
−29%
time per ticket
Same model, same real tickets, blind judges.
Dogfooding
Every change I push gets two fresh AI reviewers, never a fork of my own session, because a fork already agrees with me.
My A/B tests grade agent PRs with blind model judges, scorecards built from 1,277 real human review comments, and two ranking passes in different orders.
Models extract and judge. Plain code writes the files, checks every citation, and overrules an “approved” verdict when the build is red.
A blind test flagged a workflow step that doubled cost without improving results. Cutting it halved the cost.
Writing
Blind tests on real tickets show what happens when a coding agent gets your team’s context. More accepted changes, zero serious bugs, and half the cost.
What happens when documentation stops being a destination and becomes living context for product, engineering, and AI.
Written artifacts are just approximations; the real system lives in the team's head.
Open source
Persistent context for people and coding agents. Sources stay untouched, a nightly compile turns them into linked pages, and each agent session loads the smallest map that fits its task. In blind tests it raised accepted changes 45% and cut cost per accepted change 47%.
Recursive Language Model plugins for Claude Code and Codex. An agent reasons over sources far bigger than its context window by sending focused questions over chunks, with budgets and dry runs.
Context curation for people. A Slack assistant that ranks technical RSS with Claude and posts a short daily digest instead of a firehose.
Also: Neverending Story Pointer, free planning poker, and How Unique Is a GUID?, an interactive probability explainer.
Let's compare notes
I'm open to remote senior and staff roles in context engineering, agent platforms, and MCP infrastructure, and always up for comparing notes.