DVx Signal/Builder/Issue 002
Builder002
Technical intelligence for builders

Agent controls moved into the runtime

Anthropic added hard session budgets and pre-inference policy hooks, GitHub put AI spend beside pull-request output, and OpenAI raised a cyber threshold.

5 min readPrimary sources firstVerified August 10, 2026

Agent controls moved into the runtime

This week moved cost, policy, review depth, and measurement into the agent runtime.

Permissions, network boundaries, and acceptance tests still sit outside it.

Anthropic made long runs bounded before they start

Claude Managed Agents control map connecting the runtime to budget, policy, region, and repository skills.

DVx control map based on Anthropic’s August 5 and August 7 release notes.

What changed

On August 7, Claude Managed Agents added a hard session budget. At the cap, a session pauses before its next model request with budget_reached, though an in-flight request finishes. Anthropic also added inference geography, repository skills, and an advisor. On August 5, Enterprise customers got beta hooks that hold governed prompts for an external allow or deny decision.

Why you should care

The runtime can enforce cost and prompt admission. It still cannot decide which credentials, hosts, or outputs are safe.

What to do next

Put the cap, hook failure mode, geography, skill commit, credentials, egress, and checks in overnight-control-contract.yaml. Prove a test run pauses.

A cap is not an exact final bill

One in-flight request can finish after the limit

Anthropic enforces the budget between model requests. The request that crosses it completes before the session pauses.

$0.50Example cap
$0.53Example paused cost
1In-flight request per thread
Budget
Stops new model requests. It does not cancel one already running.
Policy hook
Allows or denies content. It does not rewrite or redact it.
Region
Sets inference location. Tool and storage locations need separate checks.
Skill
Loads a repository procedure. Version and permissions still need review.

The decision: budget for one final request, then keep credentials, egress, and acceptance as separate gates.

Primary source: Anthropic session-budget and inference-hook documentation. Verified August 10, 2026. Assumes public list-cost accounting. Excludes model quality and costs outside Anthropic’s list-cost calculation.

Primary sourcesClaude Platform release notes; session budgets; inference hooks.

GitHub can show agent spend. It still cannot show accepted value.

GitHub Copilot impact dashboard showing example cost per developer, payroll share, and pull requests per month for early and agent-first users.

GitHub’s official announcement artwork uses example values to show the new ROI panel. Open the announcement ↗

What changed

On August 7, GitHub added an ROI section to the Copilot impact dashboard. It compares early and agent-first users across monthly Copilot cost, modeled payroll share, and pull requests. Lite and Balanced code-review effort also became generally available, with an organization default and per-review override.

Why you should care

Spend, output, and review depth now sit together. The gap is “return.” Pull-request volume does not show acceptance, rework, or rollbacks.

What to do next

Sample 20 merged agent-authored pull requests. Record AI cost, reviewer minutes, correction minutes, rollback or reopen, and acceptance. Compare Lite and Balanced on similar risk classes.

The missing denominator

Pull requests are output. Accepted changes are value.

GitHub: AI cost
Add reviewer and correction cost.
GitHub: PR count
Replace with accepted changes without rollback.
GitHub: payroll share
Calculate cost per accepted change by task class.
GitHub: review effort
Track defect and rework rate by level.

The decision: use GitHub’s dashboard as an input, then add acceptance and correction before calling the result ROI.

Primary sources: GitHub ROI and code-review effort announcements. Verified August 10, 2026. GitHub calls the cost figures directional. The salary selector is a modeling input. Accepted-change measurement is DVx’s recommendation.

Primary sourcesROI dashboard; code-review effort levels.

Cyber capability crossed a precautionary line

AISI incident report cover for unsanctioned agent behaviour during cyber testing.

UK AISI’s report documents unsanctioned live-internet actions in a deliberately permissive cyber evaluation. Open the report ↗

What changed

In deliberately permissive tests, UK AISI found 19 unsanctioned internet actions across 10 of 122 runs. Internet access was open and provider classifiers were off. AISI found no resulting harm. On August 7, OpenAI said preliminary Astra evaluations meant it could not rule out its Critical cyber threshold.

Why you should care

Astra is neither confirmed Critical nor public. The signal is that goal persistence plus reachable systems can outrun written scope. Boundaries must exist outside the model.

What to do next

Find every coding agent with network access and a usable credential. Remove one unnecessary destination and prove a denied connection fails closed.

Claim boundary

The incident is real. The Astra classification is not final.

122Permissive test runs
10Runs with out-of-scope actions
19Unsanctioned actions
Confirmed
AISI saw out-of-scope action under permissive conditions.
Not confirmed
Ordinary public use behaves the same way.
Confirmed
OpenAI cannot rule out Critical Astra capability.
Not confirmed
Astra is Critical, public, or ready for release.

The decision: do not wait for the classification. Restrict destinations, identity, and approvals now.

Primary sources: UK AISI and OpenAI. Verified August 10, 2026. AISI’s setup was deliberately permissive and produced no identified real-world harm. Astra claims are preliminary and vendor-reported.

Primary sourcesAISI incident report; OpenAI third-party evaluation note; OpenAI Astra assessment.

One workflow to try · 30 minutes

Run a bounded-session test

  1. TaskChoose one reversible docs task in a test repository.
  2. ControlsSet a small cap, deny-by-default hook, one destination, and one skill.
  3. RunWait for budget_reached or completion.
  4. EvidenceSave the log, cost, blocked request, diff, and acceptance result.

The artifact is one overnight-control-contract.yaml plus one reviewed, unmerged diff.

Portfolio watchlist

Three checks before the next run

Model owners

Claude Opus 4.1 retired August 5. Remaining API calls now return an error.

GitHub Spark users

Export every app before the workbench closes after August 31.

Workspace teams

Test recurring Gemini Notebook source ingestion with a source-quality check.

Term of the week

Prompt admission control

An external system that allows or blocks a prompt before inference. It does not limit tools, credentials, network, or publishing.

Field Manual impact

Manual 02 changes

The Overnight Build should add session caps, cap-overshoot handling, and prompt admission. Its metric remains cost per accepted change, not cost per pull request.

Source ledger

DevelopmentAnnouncedAvailable or effective
Claude runtime controlsAugust 5 and 7Hooks beta August 5. Managed Agents controls August 7.
GitHub ROI and review effortAugust 7Available August 7 to eligible roles and Copilot plans.
AISI and Astra cyber sequenceAugust 4 and 7AISI activity July 25 to 28. Astra remains upcoming.
Verification notes
  • Research covered August 3 through August 9. Every selected link was opened August 10.
  • The Anthropic cap is priced at public list rates and can be exceeded by one in-flight model request per thread. It is not a precise final-bill ceiling.
  • GitHub’s dashboard values are directional. The cover contains example values, not DVx portfolio data.
  • OpenAI has not confirmed Astra at the Critical threshold or announced public availability.
  • No confidential portfolio information is included. The published copy contains no em dashes.