Software Lead Weekly 
Issue #728, 6th November 2026
This Week's Favorite
Struggle Together
4 minutes read.
Worth stealing Addy Osmani's checklist and running it on your next non-trivial agent task: write down your hypothesis before asking, ask what would have to be true for this to be wrong, and have a second model argue the opposite side. Then read the command the agent ran and its output before you read its summary. It's a cheap way to teach your team what good verification looks like. Which of these could you make a default in your team's workflow this week?
Read it later via Instapaper.
Share it via Twitter or email.
4 minutes read.
Worth stealing Addy Osmani's checklist and running it on your next non-trivial agent task: write down your hypothesis before asking, ask what would have to be true for this to be wrong, and have a second model argue the opposite side. Then read the command the agent ran and its output before you read its summary. It's a cheap way to teach your team what good verification looks like. Which of these could you make a default in your team's workflow this week?
Read it later via Instapaper.
Share it via Twitter or email.
Culture
This Is Exactly How I Feel Prompting My Agents
1 minutes read.
My humble effort to help you start the weekend with a smile.
Read it later via Instapaper.
Share it via Twitter or email.
1 minutes read.
My humble effort to help you start the weekend with a smile.
Read it later via Instapaper.
Share it via Twitter or email.
Open Source, Closed Contribution
8 minutes read.
I like the idea that for the shared cores that survive, the code is no longer the product. Avishai Ish-Shalom argues the verifiable spec is, since it's the one check not correlated with any model's blind spots (SQLite's proprietary TH3 test suite is the example). If you build products with AI, how much of your investment goes into specs and tests versus the code agents write?
Read it later via Instapaper.
Share it via Twitter or email.
8 minutes read.
I like the idea that for the shared cores that survive, the code is no longer the product. Avishai Ish-Shalom argues the verifiable spec is, since it's the one check not correlated with any model's blind spots (SQLite's proprietary TH3 test suite is the example). If you build products with AI, how much of your investment goes into specs and tests versus the code agents write?
Read it later via Instapaper.
Share it via Twitter or email.
ElevenLabs CEO and Former Palantir Deployment Strategist Gives One of the Best Explanations of What an FDE Is
3 minutes read.
"In our case, FDEs are part of the product team. They are not part of the go-to-market team. They are deeply embedded in understanding what the roadmap of the product is." -- Mati Staniszewski (ElevenLabs CEO) puts it well. I agree, and the org chart matters here: an FDE reporting into sales will optimize for the deal, while one reporting into product optimizes for what the next hundred customers need.
Read it later via Instapaper.
Share it via Twitter or email.
3 minutes read.
"In our case, FDEs are part of the product team. They are not part of the go-to-market team. They are deeply embedded in understanding what the roadmap of the product is." -- Mati Staniszewski (ElevenLabs CEO) puts it well. I agree, and the org chart matters here: an FDE reporting into sales will optimize for the deal, while one reporting into product optimizes for what the next hundred customers need.
Read it later via Instapaper.
Share it via Twitter or email.
Extensible Software in the Age of LLMs
19 minutes read.
Worth reading Jeremy Morrell's post as a checklist as your team builds anything that runs user or agent code. The five properties (cheap to run, fast cold starts, control over limits, solid isolation, safe actions) apply equally to extension platforms and agent sandboxes, and the survey of options (interpreters, V8 isolates, microVMs, WASM) helps you pick a starting point. Skim the Salesforce section first: it shows this has worked at scale since 2007.
Read it later via Instapaper.
Share it via Twitter or email.
19 minutes read.
Worth reading Jeremy Morrell's post as a checklist as your team builds anything that runs user or agent code. The five properties (cheap to run, fast cold starts, control over limits, solid isolation, safe actions) apply equally to extension platforms and agent sandboxes, and the survey of options (interpreters, V8 isolates, microVMs, WASM) helps you pick a starting point. Skim the Salesforce section first: it shows this has worked at scale since 2007.
Read it later via Instapaper.
Share it via Twitter or email.
Peopleware
Why Books Don't Work
17 minutes read.
Great take by Andy Matuschak on why explanatory books mostly fail: the medium quietly assumes that people absorb knowledge by reading sentences. Swap "book" for "onboarding doc," "architecture RFC," or "engineering handbook," and the observation stays the same. What do you do in your team today to check that the idea landed, beyond sharing the document? I'm a huge fan of Andy Matuschak and took a lot of inspiration from how he structured his blog and launched Quantum Country (look it up!).
Read it later via Instapaper.
Share it via Twitter or email.
17 minutes read.
Great take by Andy Matuschak on why explanatory books mostly fail: the medium quietly assumes that people absorb knowledge by reading sentences. Swap "book" for "onboarding doc," "architecture RFC," or "engineering handbook," and the observation stays the same. What do you do in your team today to check that the idea landed, beyond sharing the document? I'm a huge fan of Andy Matuschak and took a lot of inspiration from how he structured his blog and launched Quantum Country (look it up!).
Read it later via Instapaper.
Share it via Twitter or email.
Ditch Your to-Do List and Do This Instead (Video)
8 minutes read.
Sam Corcos on why the to-do list fails: "Your time is finite, and the number of digital items you can add to your to do list is infinite. [...] You are working with the wrong constraint, just like the amount of items you can fit in a database row, as opposed to the number of things that you can fit in your finite time of your calendar." The same trap shows up in engineering orgs: a backlog that can grow forever hides the fact that the team has a fixed capacity. If AI makes it cheaper to add work, the capacity question gets more nuanced and important to reason about.
Read it later via Instapaper.
Share it via Twitter or email.
8 minutes read.
Sam Corcos on why the to-do list fails: "Your time is finite, and the number of digital items you can add to your to do list is infinite. [...] You are working with the wrong constraint, just like the amount of items you can fit in a database row, as opposed to the number of things that you can fit in your finite time of your calendar." The same trap shows up in engineering orgs: a backlog that can grow forever hides the fact that the team has a fixed capacity. If AI makes it cheaper to add work, the capacity question gets more nuanced and important to reason about.
Read it later via Instapaper.
Share it via Twitter or email.
Harness Engineering for Self-Improvement
25 minutes read.
Lilian Weng breaks down how Agentic Harness Engineering keeps self-evolving agents honest: "Edits are only applied to the harness workspace. the runs directory, tracer, verifier, and LLM configuration are read-only, which disables a set of reward hacking (e.g disabling the verifier, swapping the model, or raising the reasoning budget) and thus it can keep every recorded gain attributable to harness edits." The evaluator and permissions live outside the loop being optimized. In a way, very similar to how you think about an organization. The playbook and rules need to serve the business more than the team or individual, yet still leverage talent to align the company's goals with teams' careers and aspirations.
Read it later via Instapaper.
Share it via Twitter or email.
25 minutes read.
Lilian Weng breaks down how Agentic Harness Engineering keeps self-evolving agents honest: "Edits are only applied to the harness workspace. the runs directory, tracer, verifier, and LLM configuration are read-only, which disables a set of reward hacking (e.g disabling the verifier, swapping the model, or raising the reasoning budget) and thus it can keep every recorded gain attributable to harness edits." The evaluator and permissions live outside the loop being optimized. In a way, very similar to how you think about an organization. The playbook and rules need to serve the business more than the team or individual, yet still leverage talent to align the company's goals with teams' careers and aspirations.
Read it later via Instapaper.
Share it via Twitter or email.
Inspiring Tweets
@nikitabier: Sometimes a demo is simply to remind the world that your competition has no desire to dream -- and no ability to evolve.
- Oren
P.S. Can you share this email? I'd love for more people to experiment and improve their company's culture.
P.S. Can you share this email? I'd love for more people to experiment and improve their company's culture.