#145 "Harness Writing Is Not Hard": Build to Learn, Not to Run with David Soria Parra // Member of Technical Staff @ Anthropic

David Soria Parra on why he'd build a custom agent harness only to understand it, not to run it,and why the industry's MCP-vs-CLI debate misses the point entirely. BROUGHT TO YOU BY: Blocks Sponsored by Blocks: Save at least 20% on your AWS costs with AI-powered optimization and enterprise discounts. Get your free Cloud Check at https://blocks.cloud/alphalist?utm\source=alphalist&utm\medium=podcast&utm\_campaign=blocks-podcast-2026 David Soria Parra co-created MCP, used to help write Mercurial, and now works alongside the people building Claude at Anthropic — so when he says building your own AI agent harness isn't actually that hard (but probably isn't worth doing anyway), it's worth sitting with. This one's a grounded, unhyped look at what to actually build versus configure. **Topics covered:** - David's nerd origin story: a PHP guestbook at 13 and a very expensive ISDN phone bill - From Mercurial core contributor to Meta's internal source control team - What Meta and Anthropic both get right by running on almost no process - David's personal setup — a handful of skills, a few MCP connectors, staying "vanilla" - Why his workflow moved from Claude Code sessions into Slack threads with Claude - The real MCP-vs-CLI debate, and why it isn't actually a debate - Should every CTO or engineer build their own agent harness? - What's still defensible in software engineering as agents get stronger - Smart-home and hardware side quests: securing a UniFi network, hacking a coffee machine - Closing time machine question, via Mercurial's own abandoned time-travel Easter egg **Timestamps:** \[00:00:51] Intro and guest background \[00:02:39] David's role in creating MCP — "one of the co-creators" \[00:03:24] David's nerd origin story — a PHP guestbook at 13 \[00:08:32] From application development to obsessing over tools and infrastructure \[00:09:38] Joining the Mercurial community \[00:12:39] How Mercurial led to Meta \[00:14:30] Biggest lessons from Meta — optimism, incentives, and minimal process \[00:20:15] Senior vs. junior engineers in the AI era \[00:22:26] Why high agency matters more than learning a narrow craft \[00:23:13] How Anthropic organizes teams in a fast-moving field \[00:25:46] Loose org process vs. strict model-safety process \[00:29:15] David's personal setup: Claude Code, a handful of skills, MCP connectors \[00:30:09] Personalized skills — writing voice and multi-agent code review \[00:32:05] Combining CLIs, MCP servers, Slack, and Linear \[00:33:00] Why his workflow moved from Claude Code sessions into Slack threads \[00:34:56] Permissions and compartmentalization for company-wide agents \[00:38:34] David's private setup — Herder, Pi, and running multiple agents \[00:45:55] Context selection, RAG, Obsidian, Git, files and folders \[00:47:42] General-purpose agents vs. custom automation — where each wins \[00:49:38] Why he'd build a harness to understand it, but not run one \[00:52:43] The new age of custom software — workout apps, home inventory, recipes \[00:53:31] Reliability, and the ongoing case for buying software \[00:55:14] How software engineering roles may shift, and what stays defensible \[00:59:38] New software possibilities for smaller organizations \[01:00:02] MCP today: the co-creator's view on his own invention \[01:02:18] The real MCP-vs-CLI debate — and why it isn't actually a debate \[01:04:27] Why building an internal MCP registry is "the better investment" \[01:06:02] What to check for when evaluating an agent-connectivity vendor \[01:06:57] Smart-home and hardware side quests — UniFi, coffee machines, and Bluetooth \[01:10:09] David's closing recommendation: stop optimizing your setup \[01:13:00] Talk to customers instead of building a software factory (unattributed exchange — see quote flag) \[01:13:38] Outro: the "hq shelf" Mercurial Easter egg \[01:14:51] What David would tell his younger self What you'll learn: - Why "harness writing is not hard" — and what actually separates a good agent harness from a mediocre one - When to reach for MCP vs. a CLI (and why the industry debate between them misses the point) - How Meta and Anthropic both run on almost no internal process — and what that requires of the people you hire - Why David moved his entire day-to-day workflow from local Claude Code sessions into Slack threads - What to check for in a vendor contract if agent connectivity matters to you - A grounded gut-check for anyone convinced they need to build a custom "software factory" Quotes: "A lot of my workflows nowadays are: 'Hey Claude, can you look into this thing, review it, figure out what you feel about it — make an adversarial review — and come back to me.' Or, 'I need to change this over here, implement this new feature, go do it, and come back when you have a review.' A lot of my interaction moved away from Claude Code sessions to threads in Slack where I interact with Claude, and Claude does the thing for me." — \[00:33:00] (cleaned of filler/false starts — verify against audio) "You don't need only MCP. You need also MCP, but you also need CLIs… Enterprises in particular love centralized things they can control via gateways in the middle. MCP is great if you need to govern the whole thing. But if you're running a local harness for your own development, running CLIs is perfectly fine. They both will be needed — they're just different for different use cases." — \[01:02:42] "Harness writing is not hard. It gives it a bunch of tools, gives it a bunch of execution steps, be a little smarter about context selection, and go for it. It turns out the model does 90% of the work." — \[00:50:00] (cleaned of stutters — verify against audio) "The main recommendation is not to over-focus on optimizing your setup. This is a very fast-moving technology — you don't need to fiddle with everything. Pick one or two things you find useful, be it Claude Code, be it Claude Tag, and just focus on using it quickly, optimizing for as little downtime as possible so you can get stuff going." — \[01:10:53] **Relevant links:** David Soria Parra on LinkedIn: https://www.linkedin.com/in/david-soria-parra-4a78b3a Model Context Protocol: https://modelcontextprotocol.io Mercurial: https://www.mercurial-scm.org Pi coding agent (by Mario Zechner): https://pi.dev Anthropic: https://www.anthropic.com Join the conversation with 600+ CTOs sharing insights on technical leadership and tech trends in the alphalist Slack. Visit alphalist.com to apply. OUR SPONSORS Sponsored by Blocks: Save at least 20% on your AWS costs with AI-powered optimization and enterprise discounts. Get your free Cloud Check at https://blocks.cloud/alphalist?utm\source=alphalist&utm\medium=podcast&utm\_campaign=blocks-podcast-2026