
Director of Engineering @Zendesk - views are my own
Melbourne
Joined April 2009
- Tweets22.4K
- Following2.6K
- Followers1.1K
- Likes4K
Darragh Kennedy retweeted
On this slow news day, we just launched Stripe evals to measure how well coding agents build real Stripe integrations.
A few cool (to me) learnings:
⢠Our plugin helped every harness/model pair we tested. All 14 configurations scored ~3ā8 percentage points higher with Stripeās recommended skills and tools.
⢠Agents searched less and wrote fewer tests with the plugin. Web searches dropped 96% and (often extraneous) test code dropped ~40%, while scores improved.
⢠We used another agent to role-play the developer. The role-player starts by sharing its imperfectly-stated goals, asks follow-up questions, and works through a few rounds of conversation with its coding agent.
⢠The last mile is hard. Agents satisfied ~78% of criteria on average, but only ~29% of runs met every required criterion.
Results and methodology: stripe.dev/evals
Darragh Kennedy retweeted
Introducing Tool Retrieval in the Gemini API āØ
With Tool Retrieval, you can scale to hundreds of tools without context bloat. Load parameter schemas on demand to preserve selection accuracy, cut tokens, and keep your prefix cache intact.
Read the blog: goo.gle/4ALQwgM
Darragh Kennedy retweeted
We spent two days trying to hack ourselves with abliterated models. The biggest risks we found, by far, were side channels into our infra, outside the application code.
Stuff like:
- The oneoff dev env we set up to test out Coder that had way too much access
- CI runners had tailnet access for no good reason
Weāve scanned our application code many times over, but there were a lot of scary risks hiding in these side doors and third party vectors. All problems that were very easy to mitigate once we were aware of them.
Many of the high-profile hacks from the past two weeks happened this way. Application code held up fine, but the company had a massive breach due to a vulnerability in a vibecoded internal tool or an unpinned CI dependency or third party js or ā¦
Youāll see us incorporating this pov into our vuln scans. (And of course weāll continue perpetually trying to hack ourselves!)
Darragh Kennedy retweeted
The first 100,000 CVEs took 18 and a half years to publish (Sep 1999 to Apr 2018). 2026 published 100,000 in 322 days.
Darragh Kennedy retweeted
Bending Spoons just bought Mailchimp for $3.68 billion. Intuit paid $12 billion for it in 2021.
Mailchimp actually got the best price of the five big names Bending Spoons has bought, measured against each oneās peak valuation:
- Miro: $17.5B peak (Jan 2022). Sold for about $1.79B. 10 cents on the dollar.
- Vimeo: $7.2B at its 2021 spin-off. Sold for $1.38B. 19 cents.
- Airtable: $11B peak (Dec 2021). Sold for about $2.25B. 20 cents.
- Eventbrite: $1.76B at its 2018 IPO. Sold for $500M. 28 cents.
- Mailchimp: $12B in 2021. Sold for $3.68B. 31 cents.
None of these companies were DOA. Theyāre big, well-known, mostly recurring businesses with millions of customers. They stopped growing, and in 2026 a big B2B company that isnāt growing sells for 2-3x revenue.
Mailchimp is guiding to 0% growth this year on about $1.26B of revenue. In 2021 it was growing 20% on $800M, which is how it got 15x.
If you run a B2B company growing under 10%, these five deals are your comps now, not your last round. Bending Spoons will pay for cash flow it can squeeze, and a reorg comes with the check.
Darragh Kennedy retweeted
More details on the AWS Emulator software factory
builder.aws.com/content/3MLGā¦
Darragh Kennedy retweeted
Building the Agent Host Protocol for sessions that continue across devices github.blog/engineering/engi⦠< from the @github team, looking for ways to get past one integration per client
Darragh Kennedy retweeted
We got a LOT of questions about how the new Copilot Fusion works under the hood.
@aashna and friends walk us through the science and engineering details that lead to this breakthrough:
youtu.be/1kpeIbdAfoQ?si=p6Wqā¦
"Just today, GitHub Copilot Fusion was ranked as the #1 coding agent on Terminal Bench scoring 5 points over Claude Code at less than half the cost."
tbench.ai
Darragh Kennedy retweeted
Everyone's using AI to write code.
But a huge additional unlock, which less people are doing, is combining the code with the context of all your live infra. Database performance. Logs. Etc.
Agents can do so much better when your code and production context are combined.
The world needs more p99 cliff diagrams, here's how you get there:
planetscale.com/blog/buildinā¦
Darragh Kennedy retweeted
QConSF highlights the general state of confusion in the industry.
Half the speakers advocate to give agents work in smallish chunks, to avoid them latching into a minor requirement and creating a lot of complexity early just for this minor bit (that could be easily added later). Fleshing out features iteratively.
Half the speakers recommend fleshing out a complete spec with all the edge cases and complexities, and giving this long deep spec to the agent to build, test and ship.
Darragh Kennedy retweeted
New blog post, with an ice-cold take on some of the last week's systems news: brooker.co.za/blog/2026/11/1ā¦
The throughput of real-world systems can vary (a lot!) based on subtle properties of the workload they're handling.
Load tests that don't match the workload are useless.
Darragh Kennedy retweeted
in case you missed it:
muse launched in australia yesterday!! š¦šŗ š¦
the early reception has been awesome, give it a shot!
Darragh Kennedy retweeted
GitHub Copilot Evals is now in public preview to help optimize Copilot for your skills, agents, and tools.
⢠Create repeatable AI behavior tests to check results and catch regressions
github.blog/changelog/2026-1ā¦
Darragh Kennedy retweeted
SOC-2 wonāt save you from this
blog.google/innovation-and-aā¦
Close. Those. Dealsā¦with the new ChatGPT for Sales experience.
It brings together our new Sales plugin and 30+ new and enhanced partner plugins to put your sales tools and customer context to workāso your team can spend less time on busywork and keep deals moving.
openai.com/index/introducingā¦
Five Bots now have full-time roles on our accounting team. We onboard, train, and manage them just like human teammates.
Here's how we built them to handle the order-to-cash process, from provisioning customer accounts to collecting payments.
x.ai/news/grok-bot-accountinā¦
Darragh Kennedy retweeted
āEverything fails, all the time." So we stopped waiting for failover.
Multipath Resilience is now GA: identical packets across physically diverse AWS backbone paths at once. First one in wins. One flag to turn on. 99.999% SLA.
Live today with @stripe and @boostmobile. go.aws/4kmJwlo
Darragh Kennedy retweeted
I havenāt lived here in awhile, but I see no other way to answer the hundreds of texts/DMs asking me if the world is, indeed, ending.
In the span of six days, weāve experienced;
11/12 ā Microsoft outage
11/13 ā Nationwide FAA ground stop
11/16 ā Major US banking disruptions
11/17 ā NSAās classified internal network goes down
We need far more analysis of why so much is breaking at once. I donāt want to get out over my skis here, and what I am hearing does not contradict the messaging that most of this is the result of screwy updates.
But stepping back, the dots are starting to form a collective picture. We are seeing:
⢠An explosion in severe and critical vulnerabilities, with AI accelerating discovery to 3X the number of vulnerabilities discovered in the same period last year. This is collectively being referred to as āvulnageddonā or āthe vulnpocalypse.ā
⢠Mean time to exploit those vulnerabilities has collapsed to NEGATIVE SEVEN DAYS. AI allows attackers to find/exploit vulns far faster than software makers have time to fix them.
⢠Powerful AI-enabled offensive tools like ARTEX have been thrown into the hands of every third tier nation state, cybercriminal and digital arsonist, which logically creates havoc.
⢠This creates enormous pressure to compress patching windows and accelerate software updates, increasing the likelihood these fixes break things in unexpected, cascading ways.
⢠Pressure to just āAI everythingā means we are rolling out automated AI tooling on top of brittle infrastructure that canāt support it.
My generation grew up in an extraordinary period of technological stability. We worked our way up the software stack, abstracting away layer after layer, until we lost a fundamental understanding of how the underlying systems actually work.
We were told to āmove fast and break things.ā VCS told us to celebrate the idea that āsoftware would eat the world.ā Now weāre telling ourselves we can simply āroll out AIā across everything, as though nothing underneath will buckle.
We never stopped to ask: SHOULD WE?
Should we be digitizing every last nook and cranny of our companies, infrastructure, healthcare, homes, cars, bodies, lives? Especially when the complexity, interdependence, and fragility we were introducing overwhelmed our ability to understand/secure/maintain the systems we built?
Nobody every asked: What if, in the name of progress, we were creating the ripest attack surface in human history?
We are now entirely dependent on software systems that almost nobody understands end to end, that increasingly few people can repair, and that attackers can now probe at machine speed. We are hearing that incident response firms, the 911 for breach victims, are backlogged six months. Even those paid on retainer lack the bandwidth to triage every incident. The cavalry is disappearing.
We are entering a period of unprecedented technological instability.
āSoftware Beirutā is an incredibly apt analogy. We are in for a period of cascading failures, whether triggered by attackers, bad updates, or our own accumulated complexity.
This, Iām sorry to say, is just the opening chapter.



























