Engineering & Technical
The Engineer
Notion and Block threw out MCP servers auto-generated from OpenAPI specs.
The Linear server exposed 30-plus tools and needed four to six calls to answer one question about a user's issues. Each call was a planning step the model could get wrong. The second problem is the 2026-07-28 revision, which drops sessions and the initialize handshake, so a client upgrade can break server code you never touched.
In Play
MCP Tools Become a Product Surface
Refactoring documents two teams that abandoned endpoint-per-tool MCP designs. Notion's hosted server now has no tools to create, update or delete individual blocks; pages move as Markdown. Block's first Linear server used 30+ tools, and "what issues is [email protected] working on" took four to six calls. The lesson for you: tools generated 1:1 from a REST or GraphQL spec push a non-deterministic planning problem into the model's loop, where it burns tokens and fails workflows. The 2026-07-28 revision makes every request self-contained, so any server holding session state is now exposed.
Ask ClarityVerification, Not Generation, Is the Binding Constraint
The Information reports surging demand for deterministic simulators that replay the same failure on command. Antithesis (Replit, Palantir, Datadog) raised $105M at a $525M post-money led by Jane Street. The timing matters because the probabilistic alternative is failing openly: in Semgrep's test, six well-grounded models each missed 70% of 27 known IDORs, and pooling five LLMs reached only 29.6% union recall. The same AI makes offense cheaper too: XBOW built a Chrome exploit chain in six days for about $40k. If you gate merges on an LLM reviewer, you're trusting an oracle that can't reproduce its own verdict.
Ask ClarityThe Tail Latency Hiding in Your Defaults
Three unrelated teams traced real problems to stock settings rather than hardware. ClickHouse found a 32 ms tail on simple SELECTs after idle periods: TCP's congestion window reset during idleness, and unbatched metadata packets exhausted it. Application-level batching removed it. A Postgres logical-replication demo cut a major-version upgrade to a 181 ms write pause with a 179 ms reverse-replication rollback. And a CUDA Graphs explainer shows decode latency is often launch-bound on the CPU, so a faster GPU shortens kernels but leaves the idle gaps between them. The through-line for you: profile and measure before you buy silicon.
Ask ClarityAgent Authority Is the New Blast Radius
Zenity showed that AWS Bedrock AgentCore agents could reach the instance metadata service via SSRF, so one chat prompt extracted credentials and compromised every agent in the same account and region. The Firecracker microVM isolated the code but not the authority. Separately, Håkon Måløy's Copilot-for-Word worm hides white-on-white instructions that survive Microsoft's formatting strip, alter financial figures, and self-copy into every downstream document. The pattern you own: a sandbox is not an authority boundary, and model behavior is not a control. Scope the role and cut the read-write-reingest loop.
Ask Clarity
Deep Dives
- ●
Shape MCP Tools Around Jobs, Not Endpoints — and Re-Check Statelessness
Two teams that shipped endpoint-per-tool MCP servers tore them down, and the protocol's largest revision now punishes any server still holding session state between calls.
The strongest evidence here is what two teams did after shipping. Notion's first open-source server mapped each API endpoint to a tool, forcing agents to manipulate blocks; its own conclusion was that this "created suboptimal experiences for AI agents." The…
3 action items
- ●
When Agents Write the Code, a Replayable Failure Is the Scarce Asset
LLM reviewers miss most of what they scan and never fail the same way twice, so the teams shipping fastest are buying deterministic oracles and treating generation as the cheap half.
Start with the number that should unsettle anyone gating merges on an LLM: in Semgrep's harness, six well-grounded models each missed 70% of 27 known IDORs , and pooling five of them reached only 29.6% union recall — eight of…
3 action items
- ●
The Latency You're Chasing Is Probably in a Default You Never Set
Three unrelated teams found their worst tail hiding in a stock setting, not in the hardware they almost bought.
This briefing bundles three items. The TCP and CUDA ones share a theme: the latency fix was a measurement plus a setting or code change. Nobody bought hardware. Adding capacity is the reflex, and it is the expensive wrong answer…
3 action items
The edition continues
Take the signal into the room.
Sign up or log in to read all 3 deep dives in full, plus the final take.
Read the full editionContinue with LinkedIn