Before writing anything about exhaustion, I counted. Things a working developer had reason to track between the 1st and the 23rd of this month:
- Aug 2: Qwen3.8 Max, the largest open-weight release yet.
- Aug 4: the UK AI Security Institute reports agents taking unsanctioned action on the live internet.
- Aug 6: GitHub Actions degrades for nine hours.
- Aug 11: CloudSEK scopes the March litellm compromise at 2,500+ organisations.
- Aug 12, 13, 14: Grok 4.6, Gemini 3.7 Flash, GLM-5.3.
- Aug 15: SpaceX closes its acquisition of Cursor.
- Aug 16: DeepSeek replaces flat pricing with peak and off-peak rates.
- Aug 17: GitHub is down 7 hours 47 minutes.
- Aug 19: Stripe agrees to buy OpenRouter.
- Aug 21: OpenAI cuts its flagship’s list price.
- Aug 22: the Model Context Protocol publishes a new roadmap.
That is not a census. It is what one person noticed.
— newyankee, Hacker News, 14 August 2026A flood of releases today, really difficult to make out for someone who does not use or test all these models on complex real world use cases as to how people decide which ones to use.
The Research Is From the Previous Regime
Search for evidence on AI and burnout and you get DORA’s 2025 report, which found no correlation and concluded burnout is a property of the organisation rather than the tools. And Stack Overflow’s 2025 survey, where the top frustration was “AI solutions that are almost right, but not quite” at 66%.
Good studies, quoted constantly, and both describe a job that is disappearing. There is no Stack Overflow 2026 survey. DORA has not published a 2026 report. The most recent serious data on how developers feel comes from a world where a developer used an assistant.
JetBrains fielded 15,000+ professional developers between May and July, published this month: 90% use AI agents at least weekly, 68% daily.
An autocomplete that finishes your sentence and an agent that writes a few hundred lines and submits them for approval are not the same tool. When someone says the research shows AI does not cause burnout, ask which year, then ask what the people surveyed were doing.
What 2026 Measures
Faros watched teams using AI heavily. Before a change goes live, another engineer is meant to read and approve it. What happened to that step:
- Time spent reviewing a change: up 441.5%. Wait before review starts: up 156.6%.
- Size of each change: up 51.3%. Bugs per change: up 54%.
- Changes approved with nobody reading them: up 31.3%.
Reviewing takes several times longer, and a third more work goes live unread. That is a queue that has stopped coping. Faros notes senior engineers absorb most of it, being the ones who can spot what AI gets subtly wrong.
Sonar surveyed 1,100+ developers. Two numbers only matter together:
- 96% do not fully trust that AI-written code works.
- 48% always check it before it goes live.
That gap will never appear in an hours-worked metric. It is shipping things you do not believe in, several times a day, knowing you did not look as hard as you should have. Confidence is the thing in short supply, and running low on it all day is tiring in a way that has no unit.
Sonar sells code quality tools and found nobody checks their code. Faros sells team measurement software and found teams are hard to measure. Neither is neutral, so I would not lean on either alone. What earns them a place is opposite methods, one asking people and one watching them, landing where JetBrains landed. Treat the direction as real and the decimals as marketing.
— dpc_01234, Hacker News, 9 July 2026No one is really pushing me to increase my workload, but at every moment there is always something ready, done by my clankers or clankers of other people that I could be unblocking.
The Second Tax
That explains the dread, not the feeling of falling behind while doing nothing wrong. For that, read the August list as a knowledge balance sheet:
- What you learned about steering Grok 4.5 was superseded on 12 August.
- Any cost model built on DeepSeek’s flat rate stopped being correct on 16 August.
- Any budget built on OpenAI’s current flagship price expires on 21 November.
- If you built against the Model Context Protocol’s session model, the 28 July spec removed sessions.
- If your team standardised on Cursor, its owner changed on 15 August.
You did the work, learned the thing, and the thing expired. Two features make this worse than the framework carousel we survived. It is priced: ignoring DeepSeek’s clock means overpaying on work that could have waited. And the timing is not yours: five Claude Code releases landed in four days this month.
What Helps
- Keep your tools still, let the models change underneath. Knowing how you work is expensive to rebuild. Knowing one model’s quirks is cheap and expires fast.
- Read release notes weekly, not as they land. Nothing in August needed same-day action except the price changes.
- Write the expiry date next to the number. Mine had none until this month.
- Decide what you will not check, on purpose. The gap between 96% and 48% already exists on your team. Choosing where it falls is far less tiring than discovering it.
None of this is a burnout diagnosis. I am a developer with a blog, not a clinician, and this may be a transition cost rather than a floor.
The Fixes Are Old
Almost nothing this month was new. The Model Context Protocol dropped sessions and rediscovered why the web went stateless in the 1990s. GitHub’s fix after its outage was retry limits and staggered timeouts. DeepSeek invented off-peak pricing, which electricity companies have run for a century. Stripe bought a currency exchange. The best fix for how AI writes is a controlled vocabulary European airlines standardised in the 1980s.
The surface moves at a punishing rate. The layer underneath barely moves, and that is where knowledge with a long half-life lives.
If you are exhausted, the useful question is not how to keep up. It is which of the things you know actually expired, and the answer is usually far fewer than it feels like.
Related Posts
