Prisma Podcast

Tyler Hogarth
  • 0.0 (0)
  • Technology

Conversations with the people building the future of software, hosted by Tyler Hogarth, Head of Engineering at Prisma. We sit down with builders, founders, and technical leaders from inside and outside Prisma to talk about the technology actually shaping how we ship software: AI, infrastructure, developer tooling, databases, and the messy engineering decisions behind them. Some guests are building the tools you use every day. Others are wrestling with the same problems you are: scaling systems, adopting AI without the hype, making infrastructure boring in the best way. The through-line is honest, technical conversation. No sales pitch, no buzzword bingo. Just people who build things talking about how and why. Prisma is a TypeScript, AI-native infrastructure platform and the creator of the most popular TypeScript ORM in the world, used by millions of developers. That gives us a front-row seat to how modern applications get built, and this podcast is where we open that conversation up to everyone. If you're a builder or a technical leader who wants to understand where technology is heading, from the engineers living it, subscribe and join us.

  1. Oct 15

    AI Builders Weekly 13 Oct 26

    Tyler Hogarth, Søren Bramer Schmidt and Tyler Benfield open with Meta's Muse Wonder and why OpenAI's Dev Day left its story muddled while Anthropic tightened its line-up. The Stack Overflow survey leads into how trust in agents grows with experience and whether code review survives 2027. Tyler Benfield explains what he still reads: how the agent verified its work, not the lines. They unpack Jev, a decision model that answers fixed questions in milliseconds with a confidence score. It ends on seat pricing, which none of them think fits a world where one seat ships tens of thousands of lines. What we discussed 1. The median coding task is solved. Anthropic keeps pushing the frontier while OpenAI ships consumer features like Dots and Sites that feel rushed. 2. Reviews have moved off the code. Check how the agent verified the change, then spend your attention on data, rollouts, performance and security. 3. Agents write too many tests. Let them, then prune to the ones that cover user-facing behaviour. A vague "could this be simpler?" cleans up more than another rule. 4. Rules should carry philosophy, not code style. A ranked value hierarchy makes the agent explain its trade-offs and push back. Onboard it like a new hire who has never seen your product. 5. Jev belongs at a decision point that needs scale or speed: classifying leads, triaging support, gating PR comments before a sandbox spins up, watching agent logs. — Hosted by Prisma x.com/prisma prisma.io — Where to find us Tyler Hogarth x.com/tylerhogarth linkedin.com/in/tylerhogarth Søren Bramer Schmidt x.com/sorenbs linkedin.com/in/sorenbs/ Tyler Benfield x.com/rtbenfield linkedin.com/in/tylerbenfield — In this episode (00:00) Intro (00:22) Muse Wonder, a third frontier model (01:16) Dev Day's confused messaging, Anthropic's tighter line-up (03:26) ChatGPT Sites, plugins and the MCP tier gate (07:14) The median task is solved (09:33) Stack Overflow: trust in AI rises with experience (11:55) Engineers sounding like managers (13:38) Are we still doing code reviews in 2027? (14:42) Spot reviews: how did the agent verify this? (16:34) Stale ADRs, data, rollouts and performance (19:10) Agents write too many tests (24:32) Delete your agents.md and refresh the rules (25:31) Rules for philosophy, proofs in the type system (28:36) Onboard the agent like a new hire (30:18) What Jev is (34:02) Confidence scores and ranked answers (36:50) Four questions before you pick Jev (37:33) A decision gate before the sandbox (40:48) Tier differentiation and seat pricing (41:55) depot.dev's simple pricing (43:28) One GitHub seat, fifty agents (45:14) Price on value, not seats (48:39) Prisma's goal: pay when you make money (50:10) Pooled tokens and where pricing ends up — References Stack Overflow 2026 Developer Survey - stackoverflow.blog/2026/10/06/the-results-of-the-2026-developer-survey-are-here/ Muse Wonder - techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder Prisma Gremlin - prisma.io/blog/gremlin-turning-open-tasks-into-pull-requests Depot - depot.dev Mastra - mastra.ai

  2. Oct 15

    AI Builders Weekly 07 Oct 26

    Tyler Hogarth, Søren Bramer Schmidt and Tyler Benfield go through OpenAI's Dev Day and what it says about where the model providers are heading. Tyler Benfield was in the room. Dots, the plugin marketplace and login with ChatGPT look like OpenAI building an operating system, while Anthropic leans into the best coding model. The second half is advice for builders on a budget. Buy a subscription, use Opus 5.5 on high, describe the outcome you want, and let the model do the rest. What we discussed 1. OpenAI is building a platform, not just a model. Dots, plugin UI embedded in ChatGPT, login with ChatGPT and enterprise commit spend that flows to partner-hosted open-weight models. Anthropic is going the other way and betting on coding. 2. Loyalty lasts as long as the better model. Prisma moved its usage from Anthropic to OpenAI a year ago and has moved it back since Fable and Opus 5.5. 3. Use the best model you can afford and stop optimising. Subscriptions, never API token prices. Opus 5.5 is the default, and Sonnet 5.5 can build 80 percent of a typical web app on a $20 plan. 4. Describe outcomes and the environment, not steps. The heavy skills and workflows built for last year's models can mostly go. Keep the docs that explain the business and the trade-offs. 5. Effort level has to die. High is the working setting. Extra high overthinks, burns usage and adds complexity nobody asked for. A few minutes of poking at the output beats more reasoning. — Hosted by Prisma x.com/prisma prisma.io — Where to find us Tyler Hogarth x.com/tylerhogarth linkedin.com/in/tylerhogarth Søren Bramer Schmidt x.com/sorenbs linkedin.com/in/sorenbs/ Tyler Benfield x.com/rtbenfield linkedin.com/in/tylerbenfield — In this episode (00:00) Intro (01:45) OpenAI Dev Day: Dots, Sol and the plugin marketplace (03:53) MCP locked behind the business tier (06:38) AI hardware, from Dots to recording devices (08:06) Login with ChatGPT and bring-your-own tokens (09:52) Enterprise commit spend on open-weight models (12:21) Model providers are starting to look like clouds (13:17) Little loyalty: Prisma's usage moved back to Anthropic (15:23) OpenAI goes consumer, Anthropic goes coding (18:22) Could ChatGPT plugins replace phone apps? (19:21) Will consumers hand tasks to agents? (22:15) Tiers and seats get in the builder's way (23:00) Pick the best model and stop optimising (25:35) Strip the heavy skills, keep the documentation (27:50) Tell the agent the outcome, not the steps (31:01) What a $20 subscription gets you (32:40) Fable for ambiguous work, Sonnet for features (35:29) Measuring the cost of intelligence (37:50) Subscriptions, never API token prices (40:37) What a harness is (42:12) Effort levels have to die (47:34) Poking at the output beats more reasoning (49:50) A Grokbot with its own email and GitHub tests the platform (52:06) Agents with payment access — References Muse Gadgets - github.com/facebookincubator/muse-gadget-sdk AgentMail - agentmail.to Artificial Analysis - artificialanalysis.ai OpenCode Go - opencode.ai/docs/go/ Prisma Compute - prisma.io/blog/launching-prisma-compute-public-beta

  3. Sep 30

    Builders Day Speaker bdougie 23 Sep 26

    Tyler Hogarth talks with Brian "bdougie" Douglas, co-founder and CEO of Paper Compute, ahead of his Builders Day talk "Dreams, Reflections, and Inceptions". Brian explains why he gives small local models roles instead of handing everything to the largest model, using his Pokémon Red speedrun as the test bed. He has labelled every session where his CTO pushed back on Go and Rust code, then trained a small model to act as that engineer's twin reviewer. The pair land on craft: what engineers lose when they stop understanding the system, and how Brian found a new craft in agents anyway. What we discussed 1. The moments an engineer pushes back on an agent are the most useful data in a session. Label them, turn them into skills, then graduate the skills into the harness, a runbook, or a trained model. 2. Agents expose gaps in process rather than create them. Teams with real documentation and tests can let agents run free-form. Teams without them find out fast. 3. The token maxing era is ending. Best model every time makes you beholden to the provider and to your weekly cap. Sonnet is enough for a well-prescribed subagent. 4. Do not trust your own pushback if you are not the expert. Brian trained a digital twin of his CTO to review infrastructure PRs before Greptile and before a human sees them. 5. Dreams are auto memory, reflections are the per-session journal, and inceptions replay a proven session to test whether a cheaper model could have done the job. — Hosted by Prisma x.com/prisma prisma.io — Where to find us Tyler Hogarth x.com/tylerhogarth linkedin.com/in/tylerhogarth Brian Douglas x.com/bdougieYO — In this episode (00:00) Intro (00:43) Labelling the moments engineers push back on agents (02:06) Pokémon Red and giving small local models roles (03:50) Skills are a stepping stone, not the goal (04:35) Personas versus raw trade-offs in the harness (05:28) Agents expose gaps in your process (07:30) Leaving the token maxing era (08:30) Fable for planning, Sonnet for subagents (10:15) Good enough beats the latest model (13:48) Encarta, fixing your own car, and losing the craft (15:30) Everyone is building their own harness (18:55) Learning infrastructure with agents as co-pilot (20:08) Every model is a book, fine-tuning is notes in the margin (21:50) The digital twin reviewer trained on the CTO's pushback (23:32) Why Astra reinvents Go libraries (26:46) Dreams, reflections and inceptions explained (28:55) Who Builders Day is for (32:22) Eggs, coffee, and what happens at a hundred users (34:20) Finding a new craft with agents — References Paper Compute - papercompute.com Builders Day 2026 - builders.day Every Agent Session Should Be a Heist - briandouglas.me/posts/2026/09/02/every-ai-agent-session-should-be-a-heist/ Dreams, Reflections, and Inceptions - papercompute.com/blog/dreams-on-paper/ Tracing. Where Jev Fits - papercompute.com/blog/from-traces-to-training-data-jev/ Jev by TypeSafe AI - typesafe.ai Greptile - greptile.com Claude Dreams - platform.claude.com/docs/en/managed-agents/dreams

  4. Sep 30

    Builders Day Speaker CJ Brewer 23 Sep 26

    Tyler Hogarth talks with CJ Brewer, co-founder of Swirls and VP of Operations at CipherStash, ahead of his Builder's Day talk on declarative agents for real operations. CJ explains why Swirls dropped MCP and a TypeScript framework for its own DSL, so agents get sub-millisecond feedback from a language server while they write. The pair debate whether tokens become a commodity or a cost worth optimising. They land on who builds for small businesses: the person at a vacuum company who found ChatGPT, or the managed service provider they already trust. What we discussed 1. AI native software means feeding agents good context and giving them early feedback loops. Swirls built its own DSL for that after MCP cost too many tokens and TypeScript gave agents too many ways to write the same thing. 2. Well-architected codebases let early stage builders contribute immediately, because agents follow patterns and are only as good as the context they get. 3. Tokens may become a commodity, but purpose-built models like Jev suggest the real question becomes which model fits which task. Cost per task still decides who can afford the software. 4. Small businesses have builders inside them, but they hand off at the point where DNS changes and MCP servers appear. Managed service providers hold the trust, and the go-to-market runs through them. 5. Humans describe the business system and own the repercussions of architecture decisions. Everything downstream of that can be declared, and CJ argues markdown files are not a system a business can rely on. — Hosted by Prisma x.com/prisma prisma.io — Where to find us Tyler Hogarth x.com/tylerhogarth linkedin.com/in/tylerhogarth CJ Brewer x.com/zcjbb — In this episode (00:00) Intro (00:43) A year of agents, and a wedding website built from scratch (02:46) Systems thinker first, engineer second (05:04) The widening gap for people early in their career (06:51) Good foundations let new builders ship on day one (07:54) What AI native software actually means (09:10) Why Swirls dropped MCP and TypeScript for its own DSL (12:45) Feed context, then get feedback before production (13:53) Are tokens a commodity or a cost to optimise (15:32) Jev and models built for one task (17:16) Taking agents to the cleaning shop down the street (19:41) The vacuum company hero who found ChatGPT (22:00) Managed service providers hold the trust (24:02) Where builders hit the threshold and call for help (25:26) Agents don't get woken up at 2am (29:19) BGP routes and the cost of a bad change (30:00) Where determinism ends and agency begins (32:44) The talk: the world is becoming declarative (34:33) Why CJ is worth hearing at Builder's Day — References Swirls - swirls.ai CipherStash - cipherstash.com Builder's Day 2026 - builders.day Jev by TypeSafe AI - typesafe.ai Eve by Vercel - vercel.com/changelog/introducing-eve-an-open-source-agent-framework

  5. Sep 23

    AI Builders Weekly: 22 Sep 2026

    Tyler Hogarth, Søren Bramer Schmidt and Tyler Benfield compare how they actually work with coding agents right now, then get into what happens when you hand one a Linux kernel. Søren still copy-pastes everything and knows exactly why the model did what it did. Tyler Hogarth has moved to T3 Code for the UI, because keeping four agent threads in your head is getting harder as the models talk more in their own jargon. Tyler Benfield does almost none of his coding on his laptop any more, and the tool that won was the one that lets the agent keep running on a remote VM after he closes the lid. The bulk of the episode is a walkthrough of a from-scratch rewrite of part of Prisma's bare metal platform. Two engineers without a systems background are compiling their own kernel and tuning Firecracker, guided by a written vision, a ranked decision framework, and permission for the agent to run its own experiments. At one point the agent cloned Firecracker, patched an address space assumption, reran the test and recommended a hardware upgrade before anyone had noticed the problem. From there: why rewrites now make sense for the next year or two, where the bottleneck moves once code is cheap, what IP is worth when anyone can rebuild Evernote, and a first look at Jev, a classifier model that answers in hundreds of milliseconds. — What we discussed 1. Remote agents are the right way to work and the tooling is not there yet. Zed halts when the laptop disconnects; T3 Code detaches over SSH and reconnects, but dropped on Søren mid-trip. 2. Vision, priority order and hard limits. An aspirational target of a million resident VMs per host steers design choices more than any measurable rule, and isolation beats readability every time. 3. Let the agent experiment. An hour of kernel flag and Firecracker configuration sweeps replaces a week of tedious benchmarking, and the result feeds straight back into the spec. 4. Code is disposable once the outcomes are written down. The Bun rewrite from Zig to Rust cost less than an engineer-year and fixed bugs along the way. 5. Once writing code compresses, code review and vendor setup look slow. Agents need safe access to systems without a human acting as the proxy for secrets. 6. Liability may be the last moat. Customer base and brand outlast code, but a giant can still fall in a year. — Hosted by Prisma x.com/prisma prisma.io — Where to find us Tyler Hogarth x.com/tylerhogarth linkedin.com/in/tylerhogarth Søren Bramer Schmidt x.com/sorenbs linkedin.com/in/sorenbs/ Tyler Benfield x.com/rtbenfield linkedin.com/in/tylerbenfield — In this episode (00:00) Builder workflows, from copy-paste to T3 Code (02:24) Why a UI still matters when threads multiply (04:28) Coding in a VM you can walk away from (06:43) Running agents from your phone while travelling (08:04) SSH, Tailscale and trusting your own environment (10:27) Building from scratch with agents as a primitive (11:56) Specs first, then the code is disposable (13:33) The Bun rewrite and what it really cost (17:27) Rewriting Prisma's bare metal layer down to the kernel (19:11) Vision, decision framework, hard limits (21:00) Agents running their own experiments (23:06) The agent that patched Firecracker itself (26:19) Two non-experts, a few months, higher quality (29:41) Understanding what the agent built (31:42) Docs written for humans, blog ideas written by the agent (34:56) Agents still can't estimate their own effort (35:57) The bottleneck moves to review and access (38:29) What is IP worth when code is cheap (41:56) Liability as a moat (45:07) Jev: a classifier model, not a chatbot (49:13) Old school ML without training your own model (50:15) Setting expectations in onboarding — References T3 Code - t3.gg Zed - zed.dev Tailscale - tailscale.com Firecracker - firecracker-microvm.github.io Bun - bun.com Doppler - doppler.com Bending Spoons - bendingspoons.com Jev by TypeSafe AI

  6. Sep 9

    AI Builders Weekly: 07 Sep 2026

    Tyler Hogarth, Søren Bramer Schmidt and Tyler Benfield work through a week of model launches: Astra, Fable 5.1 with its Mythos update, another Gemini Flash, and a cyber model from nearly everyone. The first thing the panel noticed was not capability, it was spend. Søren used a full week's quota in half an hour, because the new models spawn subagents at their own tier instead of dropping down for simpler work. That leads to the more useful question: which tasks actually need frontier intelligence, and which ones have been solved for months. Google gets a hard look. The retreat from the frontier race makes sense to everyone; the lack of movement on the products does not. From there it's open weight models running real production workloads, the gap between subscription and API pricing, and why getting agents off your laptop is still harder than it should be. — What we discussed 1. Quota disappears because every subagent runs at the parent's tier, with no downgrade for scoped work that a cheaper model would finish. 2. Google is selling compute to rivals rather than feeding its own models, while the Gemini app remains one of the weaker agents on a phone Google ships itself. 3. GLM and Kimi handle the background work: code review, incident response, ClickHouse reporting. Three months behind the frontier is enough when the task has clear success criteria. 4. A $200 subscription is a 10 to 15x discount on tokens, so open weight models only win on cost when your workload forces you onto the API. 5. Remote agent setups still break. Zed sessions die with the laptop, mobile dispatch drops connections, and the alternative is a stack simple enough to rebuild on any fresh instance. — Hosted by Prisma x.com/prisma prisma.io — Where to find us Tyler Hogarth x.com/tylerhogarth linkedin.com/in/tylerhogarth Søren Bramer Schmidt x.com/sorenbs linkedin.com/in/sorenbs/ Tyler Benfield x.com/rtbenfield linkedin.com/in/tylerbenfield — In this episode (00:00) A week of new models, and everyone's cyber model (00:54) A week's quota in half an hour (01:44) Subagents that never downgrade (02:37) What we learnt building our own multi-agent harness (03:27) Google steps out of the frontier race (04:35) Gemini, the garden bug specialist (05:33) Why Workspace still has no agent worth using (06:30) TPUs, and selling compute to your competitors (07:36) Blessed connector lists versus dropping in an MCP URL (10:01) Are the open weight Chinese models still worth it? (12:16) Where GLM is already overkill (13:34) Cheap models and weaker local currencies (14:27) Subscriptions versus API rates (16:07) Getting agents off your machine (17:29) Devin, and the hosted environments catching up (19:25) Bare metal VMs and a laptop that stops burning (20:50) SSH, dying Zed sessions, and Herdr (23:08) Simplify the stack instead: Bun and Postgres (24:47) Hosted VMs from the model providers (26:20) The multiplayer agent thread (27:57) The multi-provider outage on 3 September (29:17) Nvidia acquires Hugging Face — References Zed - zed.dev Herdr - herdr.dev Bun - bun.com Devin - devin.ai OpenRouter - openrouter.ai Hugging Face - huggingface.co Model Context Protocol - modelcontextprotocol.io Hetzner - hetzner.com Vultr - vultr.com

  7. Sep 3

    AI Builders Weekly: 31 Aug 2026

    AI model pricing is changing faster than most teams can adapt - and the real edge now comes from building around that shift, not chasing the newest headline. Tyler Benfield, Søren, and Tyler Hogarth break down why “tokens as a commodity” is becoming the winning mindset, what cheaper open-weight models like GLM Flash mean for builders, and how agent-first workflows are already replacing manual work across the stack. The conversation opens with the OpenAI and Cursor drama, but quickly turns into the bigger strategic question: what happens when model providers can cut you off, prices keep dropping, and the gap between frontier and budget models keeps shrinking? The answer, according to the trio, is to become model-agnostic, provider-agnostic, and much more intentional about how you measure cost, quality, and task fit. Chapters00:00 — AI ecosystem updates and industry shifts05:19 — How AI models reshape builder workflows15:45 — Building internal AI agents and tools23:47 — Building a better AI code reviewer32:56 — Building adaptable software systems35:27 — Giving agents the right tool access Hosted by Prisma x.com/prisma prisma.io Where to find usTyler Hogarth x.com/tylerhogarth linkedin.com/in/tylerhogarth Søren Bramer Schmidt x.com/sorenbs linkedin.com/in/sorenbs/ Tyler Benfield x.com/rtbenfield linkedin.com/in/tylerbenfield

  8. Aug 28

    AI Builders Weekly: 25 Aug 2026

    Tyler Hogarth, Søren Bramer Schmidt and Tyler Benfield talk about what agents are doing to source control and to the way teams build. They start with GitHub's monthly commits doubling to 2.9 billion, and why Cursor's Origin and Zed's DeltaDB matter more than the outage that came with them. The conversation turns to the prompt being the human input that Git has nowhere to store. It ends on why the processes we built for lazy, clever humans do not suit eager, less clever agents. — What we discussed 1. GitHub's monthly commits went from 1.4 billion in April to 2.9 billion, and source control needs reinventing rather than more capacity. 2. The human input is the prompt now, not the code, and Git has nowhere to put it. That gap is what Zed's DeltaDB goes after. 3. A written value hierarchy beats a cast of role-playing agents. Tyler Benfield's runs isolation safety, correctness, performance, operability, readability. 4. An agent is defined by the context you take away, not the experience it built up. Handoffs between agents are mostly unnecessary, with the adversarial review agent as the exception. 5. Verification is what overloads your laptop, not the agent. Test suites, type checks and browser runs are the work worth moving off-machine. — Hosted by Prisma x.com/prisma prisma.io — Where to find us Tyler Hogarth x.com/tylerhogarth linkedin.com/in/tylerhogarth Søren Bramer Schmidt x.com/sorenbs linkedin.com/in/sorenbs/ Tyler Benfield x.com/rtbenfield linkedin.com/in/tylerbenfield — In this episode (00:00) Intro (00:14) GitHub commits double in four months (02:25) Why GitHub's competitors never broke through (04:23) What Cursor's Origin is betting on (06:04) Zed's DeltaDB and prompt-aware history (07:07) The prompt is the human input now (09:25) Access control for agents in source control (10:45) How much babysitting agents still need (13:48) Recovering from an incident at 4am (15:32) On-call escalation, but for models (15:51) Daily agents reporting on ClickHouse and traces (17:14) Self-healing telemetry and agent-raised PRs (19:50) QA moving ahead of peer review (21:46) Do we need 17 agents with human names? (23:30) Replacing the product, delivery and engineering triad (25:09) The case for letting agents debate (27:38) A value hierarchy for critical code (31:15) What actually makes one agent different (33:51) Why the review agent is the exception (36:18) Software factories that improve themselves (37:56) Verification, not the agent, melts your laptop (41:47) What people are still best at — References Cursor Origin - cursor.com/origin Zed DeltaDB - zed.dev/blog/introducing-deltadb The August 17 outage, and the work ahead - github.blog/news-insights/company-news/the-august-17-outage-and-the-work-ahead/ Prisma Gremlin - prisma.io/blog/gremlin-turning-open-tasks-into-pull-requests ClickHouse - clickhouse.com Axiom - axiom.co Mastra - mastra.ai Playwright - playwrite.dev

See All (15)

About

Conversations with the people building the future of software, hosted by Tyler Hogarth, Head of Engineering at Prisma. We sit down with builders, founders, and technical leaders from inside and outside Prisma to talk about the technology actually shaping how we ship software: AI, infrastructure, developer tooling, databases, and the messy engineering decisions behind them. Some guests are building the tools you use every day. Others are wrestling with the same problems you are: scaling systems, adopting AI without the hype, making infrastructure boring in the best way. The through-line is honest, technical conversation. No sales pitch, no buzzword bingo. Just people who build things talking about how and why. Prisma is a TypeScript, AI-native infrastructure platform and the creator of the most popular TypeScript ORM in the world, used by millions of developers. That gives us a front-row seat to how modern applications get built, and this podcast is where we open that conversation up to everyone. If you're a builder or a technical leader who wants to understand where technology is heading, from the engineers living it, subscribe and join us.

Information

  • Creator
    Tyler Hogarth
  • Years Active
    2026
  • Episodes
    15
  • Rating
    Clean
  • Copyright
    © 2026 Tyler Hogarth
  • Show Website
    Prisma Podcast