Local agent sessions entered the compliance record

This week’s read

Three changes to make before your next agent run

A local Claude Code session can now enter a six-year compliance record. An agent plugin can now travel across compatible clients. Sonnet 5’s scheduled price increase is gone.

That changes retention, plugin intake, and cost forecasts.

Data and governance

Local Claude sessions entered the compliance plane

A local Claude coding session flows through an Enterprise identity boundary into a six-year compliance archive.

DVx boundary map based on Anthropic’s August 11 Compliance API update.

What changed: On August 11, Anthropic’s Compliance API started returning transcripts from local Claude Code and Cowork sessions signed into Claude Enterprise. A Compliance Access Key with read:compliance_user_data can list sessions, retrieve metadata, and fetch messages.
Why you should care: “Runs locally” no longer means “outside the record system.” The six-year default improves eDiscovery and DLP, and creates a retention and employee-notice decision.
What to do next: Export one benign test session. Record the messages, metadata, approved retention, reviewer roles, and employee notice in local-agent-retention-decision.md.

The control boundary

Retrieval improves oversight after the session

It does not approve a prompt or tool call before it runs.

Sessions: Local Cowork and Claude Code signed into Claude Enterprise.
Reviewer: A Compliance Access Key with the required read scope.
Retention: Six years by default, or a finite organization policy.
Control type: Detective retrieval after the session, not preventive approval.

The decision: treat local agent work as an Enterprise record when the user is signed into the Enterprise account.

Primary source: Anthropic Compliance API documentation. Verified August 13, 2026. Assumes an Enterprise login. Excludes personal accounts and prompt blocking.

Primary sourcesClaude Platform release notes; Compliance API content documentation.

Agentic engineering

Plugin packaging became portable. Trust did not.

A portable agent plugin package contains skills, MCP configuration, and hooks while the scanner covers only selected parts.

DVx trust map based on the Agent Plugins 1.0 announcement and Anthropic’s scanning guide.

What changed: AWS, Anysphere, Microsoft, OpenAI, and Vercel published Agent Plugins 1.0 on August 6. Google joined as a core maintainer. On August 12, GitHub shipped support across VS Code, Copilot CLI, the Copilot SDK, and the Copilot app on all plans. One package can carry skills and MCP configuration.
Why you should care: Distribution is easier. Permissions remain client-specific, and a successful scan does not make every executable part safe.
What to do next: Before installation, record source, version, skills, MCP servers, hooks, permissions, egress, scan result, and reviewer in plugin-trust-record.yaml.

Package versus policy

A portable manifest does not carry a portable trust decision

Package: Pin publisher, source, version, and commit.
Marketplace: Review managed source, update, and override policy.
Uploaded skill: Anthropic can scan it after an owner enables the beta.
MCP server or hook: Extends the package and is not scanned by Anthropic today.

The decision: reuse the package format, then repeat the trust review in every client where it runs.

Primary sources: GitHub and Anthropic. Verified August 13, 2026. Anthropic’s scanner is off by default and does not scan Copilot installs, MCP servers, or hooks.

Primary sourcesGitHub Agent Plugins 1.0 announcement; Anthropic scanning guide.

Model economics

Sonnet 5’s lower price became permanent

A shared-workload chart compares Sonnet 5 at forty cents standard, sixty cents at the cancelled September rate, and twenty cents through Batch API.

DVx cost comparison using Anthropic’s published API rates and one fixed workload.

What changed: On August 10, Anthropic cancelled Sonnet 5’s planned September 1 increase. The standard API rate stays at $2 per million input tokens and $10 per million output tokens, instead of rising to $3 and $15.
Why you should care: The same token workload now costs one-third less than the planned September rate. Issue 001’s September forecast is now obsolete.
What to do next: Update the forecast, then measure tokens and cost per accepted change on one workload. Page-based comparisons can mislead because Sonnet 5 can produce roughly 30% more tokens than models before Claude 4.7 for the same text.

Same workload · 100k input + 20k output

The cancelled increase saves $200 per 1,000 runs

Standard$2 / $10 per MTok

Input

$0.20

Output

$0.20

$0.40per run

PlannedCancelled Sep 1

Input

$0.30

Output

$0.30

$0.60cancelled

Batch$1 / $5 per MTok

Input

$0.10

Output

$0.10

$0.20per run

The decision: correct the September forecast, then use Batch only for work that can tolerate asynchronous processing.

Primary source: Anthropic pricing. Verified August 13, 2026. Assumes identical tokens on the first-party global API without caching. Excludes subscriptions, tools, agent runtime, region, partner clouds, taxes, and private discounts.

Primary sourcesClaude Platform release notes; Anthropic pricing.

One workflow to try · 30 minutes

Run a plugin trust intake

PinRecord one plugin’s source and version.
InventoryList its skills, MCP servers, and hooks.
BoundaryRecord policy, scan coverage, permissions, and egress.
TestRun one reversible task in a sandbox.

Keep one reviewed plugin-trust-record.yaml. Do not install it in production.

Portfolio watchlist

Three checks before the next run

Copilot owners

The usage report now exposes tokens per model. Reconcile one charge, then add an outcome.

Notion Worker users

Official pages disagree on the billing start date. Check the current billing UI before enabling Auto Adjust.

JetBrains teams

Copilot added memory and Ollama. Test each separately before setting a default.

Term of the week

Detective control

A control that records or reveals an action after it happens. Transcript export is detective. A prompt gate is preventive.

Field Manual impact

Manual 02 changes

The Overnight Build should add account identity, transcript retention, employee notice, and a plugin trust record. The AI Workbench should correct Sonnet 5’s rate and keep API pricing separate from subscriptions.

Source ledger

Development	Announced	Available or effective
Local-session compliance export	August 11	Enterprise beta August 11.
Portable plugin packaging and scanning	August 6 and 12	Spec August 6. Copilot support August 12. Anthropic scanner beta August 6.
Sonnet 5 price decision	August 10	Permanent rate August 10. September 1 increase cancelled.

Verification notes

Research covered August 6 through August 12. Selected links were opened and checked August 13.
Local-session retrieval applies to signed-in Enterprise sessions. The six-year default can be replaced by a finite organization policy.
Anthropic’s scanner is off by default and does not cover every executable part of a plugin stack.
The Sonnet calculation holds token volume constant. It is an API example, not a subscription or agent-runtime estimate.
The Sonnet story corrects Issue 001’s scheduled September price line rather than repeating its launch coverage.
No confidential portfolio information is included. The published copy contains no em dashes.