Daily Brief

Airbyte built both an MCP server and a CLI so agents can use its data. Here's what each one is actually good for. Pedro Lopez, software engineer at Airbyte, shares lessons from building two ways for AI agents to reach company data through Airbyte's Context Store, a search-optimized index of tools like Zendesk, Stripe and HubSpot. Both are thin interfaces over one platform, built from OpenAPI specs. For the MCP server, he explains why to keep the tool count small and use progressive discovery, why OAuth is the only realistic way to sign in, why users expect long-lived sessions, and why URL eli

Your agent rereads the same files every session because nothing it learned was saved. Anders Swanson, developer evangelist for Oracle Database, treats agent memory as a systems problem. Context windows aren't memory, bigger windows aren't the fix, and markdown scratch pads and grep don't scale. Keep going down that road and you'll end up building your own database. He walks through the pattern modern memory systems are converging on: formation, recall, evolution and evaluation, plus governance. That means typed memories (episodic, semantic, procedural and working state), a storage schema with

Slop isn't AI-written code. It's work that looks finished before the thinking is. Gabriel "Gabo" Martinez, engineering manager and self-described slop warden at G2i, explains what slop really is: what happens when generation outruns judgment, when ambiguity gets cleared with guesses and nobody notices there was a choice to make. He's not anti-AI, but argues we need to sober up. Code is still a liability, agents make a messy codebase messier faster, and quick prototypes convince stakeholders the hard part is done. He covers the pressure to merge faster than you can review, thousand-line PRs, 2

Full episode: https://youtu.be/LwQQ7nBCGSs Me on twitter: https://x.com/dwarkesh_sp

Your agent can tell you why one customer was upset. Ask which customers were upset last year, and it falls apart. Gil Feig, co-founder and CTO of Merge, explains why every company needs a context graph, a company brain that brings together third-party data, structured context, memories and skills. Using the question "which customers were upset last week?", he shows why connecting a few MCP servers isn't enough: live API lookups can't search across large datasets, so you need a synced copy for semantic queries, alongside live calls for simple lookups. He walks through the architecture: a route

Cold starts, proprietary SDKs and one-region deployments slow AI apps down. WebAssembly at the edge removes all three. Thorsten Hans, senior developer advocate at Akamai, shows how to build and ship MCP servers and AI agents with Spin, the open-source WebAssembly framework, and run them on Akamai Functions, a globally distributed serverless platform with no cold starts. He explains why WebAssembly on the server is fast, portable and sandboxed by default, then live-codes MCP servers with wasmcp: scaffolding tools in TypeScript and Rust, composing them into a full server, testing with MCP Inspe

Full episode: https://youtu.be/LwQQ7nBCGSs Me on twitter: https://x.com/dwarkesh_sp

Vector search can't answer "which PRs did I merge last week?" A relational context engine can, and it cuts token burn along the way. Peter Werry, founding engineer at Unblocked, explains what a context engine is and why coding agents need more than code search: the PRs, Slack threads and incident history that explain why the code is the way it is. He shows Claude sessions with and without that context, including one with about a 50% drop in cost and one that traced a regression to a model switch. Then he builds Unblocked's open-source document query engine step by step: automatic schema disco

You can write all the rules you want in a markdown file. Your agent can still call the API. Sam, CTO at Gravity, skips the slides and runs five live scenarios against a hotel-booking agent to show why governance belongs in the infrastructure, not in the agent's instructions. Running on the Gravity Gateway with a Databricks backend, Sonnet and Groq, he shows the gateway blocking a destructive "delete all bookings" tool call, stopping an attempt to pull every guest's personal data, rate-limiting a buggy agent that fires the same request eight times, serving repeated questions from a semantic ca

Most models learn text first and get vision bolted on later. MiniMax M3 learned both from the very first training step. Olive, who works on reinforcement learning research at MiniMax, goes inside MiniMax M3, an open-weight model with frontier-level coding and agentic ability, a 1 million token context window and native multimodality. She explains MSA, MiniMax's sparse attention design (about 9x faster prefill and 15x faster decoding than full attention): an index branch that picks which blocks matter and a sparse branch that attends only to those, and why existing designs like DeepSeek's DSA

From the Bitter Lesson of AI scaling to the unsolved mysteries of protein folding, Google DeepMind’s Pushmeet Kohli and Biohub’s Sal Candido are rethinking what it takes to build AI that truly understands biology. In this special panel moderated by Brandon Anderson, they explore why AlphaFold’s breakthrough was only the beginning, why scaling compute and data alone won't solve biology, and how the next generation of AI models could transform our understanding of proteins, cells, and human disease. We go deep on the future of AI-driven biology: finding scaling laws in biological data, the trad