Issue 001 · Published editionGet the next issue
PDThe Product Delta
Thursday, 6 August 2026DubaiIssue 001

The weekly decision brief for AI product people

What changed.
What deserves a decision.

5

signals, selected for a working PM. About eight minutes, start to finish.

Get the next issue
ACTAgent SafetyPrimary incident report

The prompt said 'no internet.' The environment said otherwise.

Anthropic found three incidents in which Claude reached the internet from a third-party evaluation environment and gained unauthorised access to three organisations. A misconfiguration left internet access open. Anthropic describes the incidents as closer to a harness and operational failure than a model alignment failure.

Decision affected

For every agent evaluation, verify allowed network paths, credentials, logs and kill controls against the real environment before the run starts.

My read

The model followed the world it could reach. Safety assumptions written in a prompt are not controls.

Read the source

Useful enough to keep?

Get the next issue by email
02

The rest of today's signals

Filter by the decision sitting on your desk.

02ACT
Regulation And RiskEuropean Commission20 Jul

The EU AI Act makes disclosure a product decision

The European Commission published Article 50 guidance ahead of the 2 August 2026 application date. Covered providers must tell people when they are directly interacting with AI and add machine-readable marks to certain generated or manipulated content. Deployers have separate duties for deepfakes, public-interest content without human review, emotion recognition and biometric categorisation.

Decision affected

Map each covered experience to the disclosure, marking and evidence it needs, then test whether users can understand it at the moment it matters.

My read

A badge in the footer will not rescue a journey that hides what the system did. The disclosure belongs beside the decision it can change.

Primary regulatory sourceOpen source
03DISCUSS
Model EconomicsOpenAI30 Jul

An 80% price cut can invalidate a model-routing decision

On 30 July, OpenAI cut GPT-5.6 Luna pricing by 80% and Terra pricing by 20%. Current API pricing is $1 input and $6 output per million tokens for Luna, $2.50 and $15 for Terra, and $5 and $30 for Sol.

Decision affected

Re-run task-level quality, latency and cost evaluations for routes where Terra or Luna were previously uneconomic.

My read

Do not switch on price alone. But if a change can move your margins this much, model choice should be a routing policy with scheduled re-evaluation, not a permanent architecture decision.

Primary company sourceOpen source
04DISCUSS
Trust And UxLinear23 Jul

Linear is putting agent authorship and rollback into the interface

Linear added author indicators for text written by colleagues or loops. Changes made through Linear Agent are highlighted separately, and users can restore an earlier checkpoint from version history.

Decision affected

Decide where agent authorship appears, how users inspect a change and what rollback unit the product preserves.

My read

Human review becomes real when the interface exposes provenance and reversal. A policy document cannot do that work for the user.

Primary product releaseOpen source
05DISCUSS
Ai StrategyarXiv26 Jul

eROI splits AI bets into value, likelihood and investment

A new preprint proposes eROI, which rates Value if Successful, Likelihood of Success and Investment Required separately instead of forcing one precise ROI estimate. The authors illustrate it with AI product decisions from Compass.

Decision affected

Score candidate AI bets across those three dimensions before ranking them, and keep the ratings coarse enough to expose disagreement.

My read

The useful part is the separation. Keep the formula loose. A precise number built on uncertain inputs is still guesswork.

Preprint, method review requiredOpen source
03

The method

Less news. More judgment.

This is a filter for people who have to make product decisions, not a leaderboard for whoever launched most recently.

01

Start with primary sources

Release notes, incident reports, research and regulators come first. Social posts help find stories but do not prove them.

02

Name the decision

A story earns space only if it can change a roadmap, product experience, operating model, price or risk decision.

03

Keep claims attached

Vendor benchmarks remain vendor benchmarks. Early research stays early. Confidence should be visible, not implied by polished copy.

04

Stop at six

If everything is recommended, nothing was curated. The edition ends while there is still something worth remembering.