Tech Radar — Autonomy by default

The thread this week is autonomy becoming the default setting, not the opt-in. Vendors are flipping agents to run without per-step approval, pricing their output per token, watermarking what they produce, and — on the other side of the fence — automating attacks faster than most teams can patch. The control plane around AI, not the AI itself, is where leadership attention should go this week.

Claude Code flips auto mode on by default

Starting August 14, auto mode becomes the default for Claude Code on Pro, Max and Team accounts: the agent proceeds without per-step approval unless an action is judged irreversible, destructive, or aimed outside your environment (techcrunch.com). Anthropic's argument is data: in a study with 1,053 testers, automated policy caught 89% of harmful actions versus 13.6% for manual review — because humans approve 97% of permission prompts anyway.

Impact for companies: every team using Claude Code inherits the new default in days, along with new safeguards like prompt-injection screening and customizable deny rules.

Risks and opportunities: the risk is inheriting autonomy without ever having written a policy. The opportunity is admitting what the data shows — approval fatigue is real, and policy-based control scales where click-to-approve does not.

My take: defaults are decisions someone else made for you. Before August 14, review your deny rules and treat agent permissions like production IAM: explicit, versioned, audited. The teams that never configured anything are the ones this change actually affects.

Meta enters the terminal-agent race

Meta released Muse Code, a terminal coding agent powered by Muse Spark 1.2, a model co-trained with the agent harness (research.meta.ai). The design choices matter more than the launch: persistent background agents, and a local append-only event log that makes runs replay-exact and restart-safe. In Meta's case study the agent ran over 1,000 tool calls across 24 hours optimizing GPU kernels.

Impact for companies: a third hyperscaler in the terminal-agent market means real price and capability competition against Claude Code and Codex.

Risks and opportunities: the risk is workflow lock-in to any single vendor's harness while the market is this fluid. The opportunity is leverage — and a higher bar: crash-safe, auditable long runs should now be a procurement requirement, not a bonus.

My take: the interesting part is not the model, it is the event log. Multi-hour autonomous runs are only operable if you can replay and audit them — the same lesson infrastructure teams learned with event sourcing. Evaluate agents like you evaluate databases: by their failure and recovery story.

GitHub puts AI spend next to payroll

GitHub's Copilot impact dashboard gained a return-on-investment section: cost per developer per month derived from actual AI credit consumption, that cost as a percentage of payroll, and pull requests per developer per month — comparing chat-centric developers against agent-first ones (github.blog). GitHub itself labels the figures "directional."

Impact for companies: the first first-party tool that answers the question CFOs have been asking for two years, with numbers pulled from real usage.

Risks and opportunities: the risk is obvious — PRs per developer becoming a productivity KPI is how you get many small, low-value PRs. The opportunity is framing the ROI narrative before finance frames it for you.

My take: measure before your CFO does. Pair the dashboard's PR counts with lead time, change-failure rate and rework, and present the package yourself. A directional metric without context becomes a target; with context it becomes a budget defense.

OpenCost learns to price a token

OpenCost 1.121.0 integrated with llm-d to attribute GPU and infrastructure spend to models and tokens, with new metrics like cost per million tokens and a clean split between allocation-based cost (keeping the model warm) and usage-based cost (cncf.io). The worked example is the whole story: a self-hosted model costing $1.00 per million tokens compute-only is $4.00 all-in at 25% utilization — against a $2.00 external API, break-even sits near 50% utilization.

Impact for companies: platform and FinOps teams finally get an open-source answer to "what does a token cost us," validated on a 109-GPU cluster running 30 models.

Risks and opportunities: the risk is that compute-only estimates flatter self-hosting and drive premature GPU purchases. The opportunity is making build-vs-buy inference decisions with all-in numbers.

My take: this is FinOps discipline arriving at AI, and most self-hosting business cases will die at the utilization line. Measure before you buy GPUs — renting looks expensive per token until you price an idle cluster.

Anthropic watermarks Claude's output

Anthropic confirmed that models released after August 2 automatically watermark generated text and files, complying with the EU AI Act's Transparency Code (techcrunch.com). The watermark is embedded in the text itself, survives copy-paste, and applies across the API, Claude Code and other surfaces. Google, Meta, Microsoft and OpenAI committed to the same code.

Impact for companies: AI-generated code, documents and deliverables produced by your teams become machine-identifiable by third parties, regardless of which tool produced them.

Risks and opportunities: the risk sits in client contracts and deliverables that are silent about AI assistance. The opportunity is turning provenance into a feature — audit trails for regulated industries just got easier.

My take: assume everything AI-assisted is detectable and update your policies and client agreements accordingly. Chasing watermark removal is the wrong energy; being transparent about how your team works, with quality gates to back it up, is the durable position.

The AI security arms race got concrete

Black Hat USA delivered the proof points: PortSwigger demoed an autonomous system that generated genuinely novel HTTP attack classes and earned bug bounties against production systems, and Tencent showed an LLM pipeline that found over 100 logic vulnerabilities in Chrome and Android (msn.com). CrowdStrike's data: 88% of attacks exploiting public PoC code began within 48 hours of release. Days later, OpenAI restructured its cyber defense service into Blue and Red tiers, the latter with a purpose-trained GPT-5.6-Cyber model for trusted partners (techcrunch.com).

Impact for companies: remediation windows measured in days are obsolete, and frontier labs are now selling the defense to match the offense they warn about.

Risks and opportunities: the risk is access asymmetry — the strongest defensive models are gated to partner lists. The opportunity for lean teams is AI-native triage without a 24/7 SOC headcount.

My take: patch latency from PoC publication is now a metric worth reporting to the board. If your dependency-update and emergency-patch path takes a week, the 48-hour number says you are structurally exposed — fix the pipeline before shopping for AI defense products.

Layoffs: 2026 already passed all of 2025

Per Layoffs.fyi data, 125,759 tech employees were laid off across 264 companies by August 6 — exceeding 2025's full-year total (ibtimes.co.uk). The first week of August alone: Zillow cut 500+ roles despite 18% revenue growth, Etsy cut ~220 mostly in Product and Engineering, and both explicitly stated AI was not the driver.

Impact for companies: cost discipline is now targeting product and engineering organizations at growing companies — revenue growth no longer shields headcount.

Risks and opportunities: the risk for leaders is assuming a good P&L protects the org chart. The opportunity, uncomfortable as it is: teams that can show cost-per-outcome survive reviews that teams with only velocity metrics do not.

My take: "not driven by AI" is doing heavy lifting in these announcements — AI capex is squeezing opex budgets even where AI is not replacing the work. Connect your engineering spend to business outcomes before someone in finance does it for you, with less context.

The trend to watch

Autonomy is being priced, watermarked, audited and weaponized — all in the same week. The differentiator for the next quarters is not which agent you adopt, but whether you built the control plane around it: permission policies, cost attribution, provenance handling, and a patch pipeline fast enough for a 48-hour world. That is unglamorous work, and it is exactly where technology leadership earns its keep.