Dustcrawl
Agentic Digest
Agentic Digest · · Issue #84
An agent config file is a pace signal. A Markdown workflow is still an Actions job. A token count is not the customer.
Dictionary
/ˈɪnˌpʊt ˈmɛtrɪk/ · n.
A count of tokens, tool adoption, or code changes used as a stand-in for delivery. It can move with the work for a while. Aimed at, it becomes the work.
Example The slide said eight times the commits. The only new file was an agent config. Nobody asked whether a customer could tell. That chart was an input-metric wearing a badge.
Wire, filtered
GitHub's Octoverse 2026, published 28 October, is a volume report with an agent subplot. Developers pushed code 17.9 billion times in the past 12 months, more than in 2024 and 2025 combined. Commits across public and private repos reached 34.3 billion, more than three times where they stood two years ago. Merged pull requests more than tripled to 1.3 billion over that same comparison, about 3.5 million a day. Actions ran a job 1,440 times a second.
The agent lines are narrower than the headline. Among repositories created this year, the average repo with an agent configuration file logged eight times as many commits as the average new repo without one. That is file presence, not a census of who typed each commit. As of September, developers start more than 40% of pull requests with an agent, up from 17% in March. GitHub's own next sentence is the useful one: review is how developers shape what comes back. In September, 61% of Copilot users used more than one model. A start rate and a model mix are not a defect rate.
Do not mix windows. Willison, at Universe, recorded the monthly rates Jay Parikh highlighted: merged pull requests to 220 million a month, commits to 7.38 billion a month. Those are not the written report's twelve-month totals. He also quotes the methodology: country push ranks are adjusted by an LLM reading profile text, to offset VPN noise, and industry buckets come from a model labeling random samples of name, description, topics, and README. The report itself says AI mention and agent-opened pull requests do not move together. In several sectors, more than four in ten public repos mention AI while agent-opened pull requests stay under 3%. A rank built by a classifier is a rank. It is not a headcount.
GitHub Agentic Workflows is generally available as of 28 October, on github.com for Free, Team, and Enterprise. You describe a reasoning job in Markdown. The project compiles that file into ordinary Actions YAML, so the run still sits in your runner groups and policy constraints. Copilot can power it. So can another coding agent. The default Copilot CLI engine needs a Copilot license. General availability is a support label. It is not a new permission.
The changelog's safeguard
list is the part worth
copying. Agents read GitHub
content through an
integrity filter,
run read-only by default,
and execute in a sandboxed
container behind the
Agent Workflow Firewall.
Proposed changes go through
safe outputs,
then a threat-detection job
scans them before they are
applied. Since preview, the
project added self-hosted
ARC runners, a
non-experimental
gh aw forecast
that includes threat-detection
credits, per-workflow model
choice, a model router, and
micro-VM isolation with
enclaves for cross-repo
runs. None of that is the
Markdown.
The same note includes a customer quote, not a product benchmark. Ben Alton at Accenture reports a lead saving about 29 hours a week, velocity more than doubled, AI authoring 91% of that team's commits, and production errors down from about 11,000 a day to under 1,000. Treat that as one team's account. Your fence is still read-only, the sandbox, safe outputs, and the scan that runs before apply. If those are off, GA did not turn them on.
Amazon engineering leadership rewrote the software-engineer role guideline and published the memo. Terk Terkowitz, VP and Distinguished Engineer for Worldwide Amazon Stores, says the argument inside the company was whether AI is different enough to put in the role text. They decided the durable part is not the tool. It is impact on customers, the business, and the team. The memo is explicit that future practice will have AI at the center, and equally explicit that a definition of "good AI practice" written this year will look old next year.
The measurement paragraph is the one to keep. Token use, tool adoption, code changes, and reviews are input metrics. Amazon says they have correlated with the outputs that matter, partly because teams that adopted hardest delivered more, and partly because teams that did not fell behind. It also says those metrics are deeply flawed as targets. Aiming at tokens fights frugality. Aiming at lines of code fights simplifying. The suggested look is the 10th percentile (not using the tools, and why) and the 90th (using them hard, and whether delivery moved with the use).
The sentence Willison pulled out is the practice: if the current tools do not help a piece of work, that is a finding, not a failing. Write why, share it, and keep looking. Sufficient understanding, in the memo's words, means you have used the tools on production work, can name a task they helped, and can name a task they did not. A dashboard that only goes up is not that understanding.
One thing to practice
A big agent number is useful the way a speedometer is useful. It tells you something moved. It does not tell you the move was the one you wanted, or that copying the costume will reproduce it. Octoverse, Agentic Workflows, and Amazon's memo are three ways of saying that in one week.
Separate presence from authorship. "Repos with an agent configuration file did eight times the commits" is a comparison of averages among new repos. The file was present. The report does not say the agent wrote each commit, that review got easier, or that defects fell. Teams who were already shipping may be the ones who added the file. If you add the file to chase the multiple, you have optimized the classifier, not the product.
Separate start from land. "More than 40% of pull requests start with an agent" is a first-move statistic. GitHub puts review after that sentence on purpose. A draft rate can rise while a human still decides the merge, or while nobody does. Ask which. The same caution applies to windows. A monthly rate from a keynote and a twelve-month total from the written report are not the same bar. Put the window in the label before you put the number in a slide.
Separate the prompt from the fence. Agentic Workflows compiles Markdown into Actions YAML. That is a convenience. The controls that matter are outside the prose: read-only by default, a sandbox and firewall, safe outputs, and a threat-detection job before apply. General availability did not move those into the model. If a workflow can widen its own permissions, the Markdown is the asker and the asker is also the gate. That is the failure mode, not the feature.
Then refuse to aim at the correlate. Amazon's memo is willing to watch tokens and adoption because, for now, they move with delivery. It is not willing to make them the goal. A target that is only an input grows the input. The replacement is a pair: one pace number, one outcome you can explain without the pace number. And a written miss. If the tool does not help a task, the paragraph that says why is the eval. Leaders in the memo are expected to have used the tools on production work and to be able to name both a help and a miss. That is a smaller bar than "find every place an agent could have typed."
Watch the speedometer. Do not steer by it. The number that cannot name a customer, a defect, or a fence is an input-metric, and it will happily become the job.
The notes this sits on are GitHub's Octoverse 2026, Willison's Universe notes, the Agentic Workflows changelog, and Amazon's role-guideline memo.
From: Eight-times (repos: twelve, new file: agent config, commits since: zero, customer: unconsulted, slide: already titled)
Octoverse said new repos with an agent config file do eight times the commits. I added the file to twelve repos before lunch. Can I put 8× on the status slide and call the migration done?
A: No. The eight times is an average of new repos that already had the file, compared with new repos that did not. It is not a coupon. Those repos may also be the ones whose owners were already shipping. A config file with no runs is a costume. Count commits if you want a pace number, and label the window. Count a defect that stopped escaping, or a customer task that got easier, if you want an outcome. Adding the file to hit the chart is the chart using you.
From: Tokenmaxxer (goal: tokens per week, direction: up, outcome: TBD, memo: printed, the careful paragraph: skipped)
Amazon said input metrics correlate with the outputs that matter. I set the team's goal to tokens per week. They even invented a word for going too far, which means they thought about it. Isn't aiming at the correlate how you get the outcome? The workflow is generally available, so the Markdown can approve its own writes now, right?
A: Read the next sentence. They call those metrics deeply flawed as targets, because a target that is only an input grows the input. Tokenmaxxing fights frugality. Counting lines fights simplifying. Correlation is a reason to watch the 10th and 90th percentiles and then go look at the work. It is not a reason to make tokens the goal. And no, general availability did not hand the workflow a pen. The default is still read-only. Safe outputs and a threat-detection job sit in front of apply. If your Markdown can widen that, you built the bug. Bring the paragraph about the task the tool did not help. Leave the bigger token number in the drawer.