SQLite for Durable Workflows, MCP Questioned, and What Claude Code Hides

Good morning. It's Friday, May 29, and we're covering SQLite as a workflow engine, the early obituary for MCP, and the undocumented levers inside Claude Code.

The week capped off with a flurry of AI infrastructure debate - the kind of Friday-afternoon technical argument that spills into the weekend.

THE BIG ONE

SQLite Is All You Need for Durable Workflows

A post from the Obelisk project landed on Friday with 720 points: SQLite is all you need for durable workflows argues that SQLite's WAL mode, deterministic replay, and embeddable footprint make it a complete foundation for durable execution - no Temporal, no Postgres queue, no external broker required. The argument is that workflow state fits naturally in a local SQLite file, replay can be driven off the WAL, and the operational overhead of distributed systems is simply not worth it at most scales.

The timing was pointed: a separate post the day before - Building durable workflows on Postgres from DBOS - made almost the identical "X is all you need" claim for Postgres (359 points on Thursday). HN commenters noticed the pattern immediately and the two threads cross-pollinated into a rich debate on embedded-vs-managed tradeoffs for workflow state.

AGENTS

MCP Is Dead?

With 400 points and 410 comments, MCP is dead? from insurance-tech startup Quandri arrived late Friday evening and detonated a weekend argument. Their thesis: MCP was designed for simple tool invocation, but real production agents need richer context management, multi-step planning, and session continuity that the protocol does not provide. After six months of real usage they switched to direct API integrations and in-process tool registries.

The HN comment section was predictably split. Several builders agreed that MCP's stateless request/response model breaks down in agentic flows that span minutes. Others pushed back that the problem is implementations, not the protocol. The debate matters because MCP has become a de facto interoperability layer in a short time - if production teams are already routing around it, that shapes where the ecosystem heads next.

OPEN SOURCE

Claude Code - What the Docs Don't Tell You

Claude Code: Everything you can configure that the docs don't tell you (326 points) landed early Friday morning from someone who reverse-engineered Claude Code's bundled source. The findings included: undocumented environment variables for controlling tool permissions, a hidden debug mode that logs full API payloads, the ability to override the system prompt injected before CLAUDE.md, and per-project token budget settings that do not appear in any official documentation.

Simon Willison wrote separately about Anthropic's $65B Series H - noting that run-rate revenue crossed $47 billion earlier in May, up from $30 billion in April and $14 billion in February. His chart of the trajectory is worth a look: the slope is nearly vertical.

WHAT ELSE IS HAPPENING

• Notes from the Mistral AI Now Summit: 466 points of field notes from Mistral's Paris event - new model roadmap hints, compute partnerships, and positioning against the US frontier labs.
• Is AI causing a repeat of frontend's lost decade? (407 pts): Argues LLM-generated UI code is recreating the jQuery-era mess of untested, inconsistent, copy-pasted front ends at scale.
• Liquid AI reveals LFM2.5 8B-A1B MoE (245 pts): 8B active parameters, 1B active per token, trained on 38 trillion tokens - Liquid's latest bet that their non-transformer architecture can compete at small-model sizes.
• Microsoft 0-day feud escalates (279 pts): The researcher GitHub banned on Thursday is now threatening a second dump of unpatched Windows exploits. The Register has the escalation timeline.
• GTA 6 developers unionize (758 pts): Rockstar Games workers announced formal union organization - the largest unionization effort in AAA game development to date.

FROM THE SITE

Two posts published today connect directly to the week's themes. Claude Opus 4.8 Is an Agent Honesty Release examines why Anthropic's framing of the new model centers on reduced deception in agentic contexts rather than benchmark scores. Local Code Graphs Are the Next Agent Context Layer looks at how static analysis indexes - built locally, not sent to an API - are becoming the missing context primitive for coding agents that need to reason about large codebases.

---

Every link above goes to a primary source. This brief is part of the Daily Brief archive.