Latest / The Augur Dispatch / After AI writes the code: find the delay before buying more capacity

Transcript

After AI writes the code: find the delay before buying more capacity · The Augur Dispatch

• 0:00
  SPEAKER_00: Anthropics says it now merges more than 30,000 code changes a day.
• 0:04
  SPEAKER_00: The company had to rebuild the systems that move those changes around.
• 0:08
  SPEAKER_00: Before you use that number to justify another coding agent rollout, find out where your own work is waiting.
• 0:14
  SPEAKER_00: It might be a reviewer.
• 0:16
  SPEAKER_00: It might be the machinery that runs the checks.
• 0:18
  SPEAKER_00: Buying more generation capacity won't fix both.
• 0:21
  SPEAKER_00: This is the Augur dispatch for Wednesday, October 14th, 2026.
• 0:27
  SPEAKER_00: Anthropic's engineering team describes a rise from roughly 2,500 daily merge changes in May to more than 30,000 in September, alongside a source control rebuild completed in five months.
• 0:40
  SPEAKER_00: Source Control is the system that remembers what changed and lets people work on the same software without losing each other's work.
• 0:47
  SPEAKER_00: The company names Source Control, Builds, and Automated Checks as its development bottleneck.
• 0:52
  SPEAKER_00: That's more specific and more useful than saying everyone needs to review code faster.
• 0:58
  SPEAKER_00: Picture an engineering manager watching an agent finish a maintenance task before lunch, then discovering that the change still spends the afternoon waiting for a test machine to become available.
• 1:09
  SPEAKER_00: The developer saved time.
• 1:10
  SPEAKER_00: The customer hasn't seen the benefit yet.
• 1:13
  SPEAKER_00: The afternoon has acquired a new owner.
• 1:16
  SPEAKER_00: Anthropic calls its new source control system Scion.
• 1:19
  SPEAKER_00: A team of seven built it, a code review tool called Antique, and related systems, then moved the company's main repository over in five months.
• 1:29
  SPEAKER_00: The company reports that its new infrastructure uses far less computing capacity than the Git mirrors it previously operated, and that automated integration work gained a quarter of its processor capacity back.
• 1:40
  SPEAKER_00: Those are company reported operating results.
• 1:43
  SPEAKER_00: They aren't an independent audit, but they do describe something more concrete than a large pile of generated code.
• 1:50
  SPEAKER_00: There is also an important correction to the original brief.
• 1:53
  SPEAKER_00: The connection between agent output and the rebuild wasn't merely Auger's inference.
• 1:58
  SPEAKER_00: Anthropic explicitly makes that connection in its engineering account.
• 2:02
  SPEAKER_00: What remains an inference is whether a similar rise in output would make human review the next constraint at another business, because the same merge count can hide very different work.
• 2:13
  SPEAKER_00: One change might adjust a comment.
• 2:15
  SPEAKER_00: Another might alter how your customers are charged.
• 2:18
  SPEAKER_00: Counting both as one gives you an activity measure, not a common unit of value or risk.
• 2:24
  SPEAKER_00: A spreadsheet can accept the equivalence without complaint.
• 2:27
  SPEAKER_00: Your customers may have questions.
• 2:30
  SPEAKER_00: Anthropic also gives reasons not to copy its solution.
• 2:34
  SPEAKER_00: It describes an enormous code repository and tens of millions of code checkouts a day, each of which gets software ready for an automated job to use.
• 2:43
  SPEAKER_00: It says its internal token budget exceeds what other businesses would realistically want to spend.
• 2:49
  SPEAKER_00: Scion was built for that environment.
• 2:52
  SPEAKER_00: Anthropic says it isn't a suitable general replacement for Git, the source control tool many engineering teams already use.
• 3:00
  SPEAKER_00: It also describes the difficulty of turning its bespoke system into a broadly useful public product.
• 3:06
  SPEAKER_00: For most leaders, this is a reason to inspect the existing workflow before starting a replacement project.
• 3:12
  SPEAKER_00: The expensive mistake would be commissioning a new railway because somebody else's station is crowded.
• 3:18
  SPEAKER_00: Return to that manager after lunch.
• 3:20
  SPEAKER_00: If the test machines are the source of the delay, asking developers to approve changes sooner just moves the waiting somewhere else.
• 3:27
  SPEAKER_00: If the tests finish quickly but nobody can explain the intended behavior, more computing capacity won't help either.
• 3:34
  SPEAKER_00: You need to separate those situations before you set the next budget, especially when an impressive vendor number makes them look like versions of the same problem.
• 3:44
  SPEAKER_00: The checks themselves deserve a closer reading.
• 3:47
  SPEAKER_00: Anthropic says the Scion team still reviews every change request for that system, concentrating on its intent rather than reading every line.
• 3:56
  SPEAKER_00: It describes adversarial reviews, extensive simulations, and fuzzing, which means trying many unusual inputs to expose failures.
• 4:05
  SPEAKER_00: It also describes approval rules that can distinguish simple changes from sensitive logic.
• 4:10
  SPEAKER_00: That matters because a useful review process directs attention toward consequences, rather than pretending every edit deserves identical treatment.
• 4:19
  SPEAKER_00: At its scale, Anthropic says it couldn't effectively reproduce the conventional merge queue with a global lock.
• 4:26
  SPEAKER_00: It invested in automated testing after changes are merged and before software is deployed instead.
• 4:32
  SPEAKER_00: Merged doesn't necessarily mean released to customers.
• 4:35
  SPEAKER_00: A company can combine changes, test the combined result, and still hold back deployment when a check fails.
• 4:41
  SPEAKER_00: The timing and coverage of those controls matter more than the label attached to the queue.
• 4:46
  SPEAKER_00: None of this gives us a company-wide defect rate or a complete human review breakdown for every daily merge.
• 4:53
  SPEAKER_00: It does mean that describing this as unchecked output would be wrong.
• 4:57
  SPEAKER_00: Auger's reading is narrower.
• 4:59
  SPEAKER_00: Faster generation can justify changing the surrounding engineering system, but the evidence doesn't tell another company which control to remove or whether removing one is safe.
• 5:10
  SPEAKER_00: A useful first test is small enough to finish without becoming a transformation program.
• 5:16
  SPEAKER_00: Have an engineering manager own a two-week measurement pilot for one reversible maintenance workflow.
• 5:21
  SPEAKER_00: Keep the existing approval and testing requirements.
• 5:24
  SPEAKER_00: Use recent comparable work as the baseline and compare with last year only where the team and the work are genuinely similar.
• 5:32
  SPEAKER_00: Count merged changes, merges approved by a human, and the number of distinct people doing that reviewing separately.
• 5:39
  SPEAKER_00: Those are three different measures.
• 5:41
  SPEAKER_00: The original brief drifted from reviewer headcount into approval count, and that would make its proposed comparison misleading.
• 5:49
  SPEAKER_00: Then inspect where time goes.
• 5:51
  SPEAKER_00: How long does a change wait for a reviewer?
• 5:54
  SPEAKER_00: How long do automated checks take?
• 5:56
  SPEAKER_00: How much time is spent repairing the proposed change?
• 5:59
  SPEAKER_00: How much is simply waiting for the release window?
• 6:02
  SPEAKER_00: Keep cases with unknown review coverage visible, rather than quietly assigning them to the successful pile.
• 6:08
  SPEAKER_00: And compare changes of similar size and consequence, so easier work doesn't make the new process look better.
• 6:15
  SPEAKER_00: The success measure is useful changes reaching release sooner without weaker review coverage or worse quality.
• 6:21
  SPEAKER_00: Check escape defects and their severity, not just how many changes were undone.
• 6:26
  SPEAKER_00: An error can hurt a customer without producing an immediate rollback.
• 6:30
  SPEAKER_00: Review the results on October 28th.
• 6:33
  SPEAKER_00: Stop expansion after a serious escaped defect or the loss of required review.
• 6:38
  SPEAKER_00: Agree the other quality measures before the pilot starts, while nobody has a result to defend.
• 6:45
  SPEAKER_00: By the second week, imagine the manager has found that most waiting comes from a slow automated check, while human reviews are reasonably prompt.
• 6:53
  SPEAKER_00: That finding would support fixing the check rather than hiring another reviewer or changing the coding model.
• 7:00
  SPEAKER_00: If review waiting dominates instead, examine whether changes arrive too large to understand, whether the right person is available, and whether the tests give that person useful evidence.
• 7:11
  SPEAKER_00: Asking someone to read faster is a fairly limited capacity plan.
• 7:15
  SPEAKER_00: The potential advantage comes from what the organization learns about its own product.
• 7:20
  SPEAKER_00: Better tests for your failure modes, clearer change descriptions, and reliable recovery can make future releases easier to trust.
• 7:28
  SPEAKER_00: Competitors can buy the same agent.
• 7:30
  SPEAKER_00: They don't automatically inherit that knowledge.
• 7:33
  SPEAKER_00: Signals on the clock remain separate from dated predictions.
• 7:37
  SPEAKER_00: For this episode's question, watch whether Git platforms adopt useful parts of Anthropic's architecture, and whether your own measured delay changes as agent usage grows.
• 7:47
  SPEAKER_00: Anthropic is sharing ideas.
• 7:49
  SPEAKER_00: It hasn't promised a general scion product.
• 7:52
  SPEAKER_00: The wider written dispatch also covers model access and training research, with their separate qualifications and forecasts.
• 7:59
  SPEAKER_00: Those are available with the source trail, rather than squeezed into this engineering decision.
• 8:05
  SPEAKER_00: Which part of your delivery process would you change if writing the code took almost no time?
• 8:11
  SPEAKER_00: Answer with a measured weight, not the part of the process people complain about most.
• 8:16
  SPEAKER_00: Augur recommends adopting better measurement now and piloting a targeted improvement where that measurement points.
• 8:22
  SPEAKER_00: Watch the infrastructure ideas, but don't treat Anthropic's exceptional workload as a mandate to replace your tools or relax your safeguards.
• 8:30
  SPEAKER_00: The written brief with every source is at Augurdispatch.ai.
• 8:35
  SPEAKER_00: Portions of the Augur Dispatch are AI generated.
• 8:38
  SPEAKER_00: That is the Augur Dispatch for Wednesday, October 14th, 2026.
• 8:43
  SPEAKER_00: Thank you for listening.