Structural ratchets for AI-assisted Go, six months into make verify

The agent found every page certified unreachable. Links cut search cost and keep working when search is off; they do not change what it reaches

A wrong note that passed review spread only where the agent could have checked it, and only on Haiku 4.5: 16 of 24, against 0 of 24 on Sonnet 5 and Opus 5

Adding your own GPUs to the platform with the NVIDIA GPU Operator

The harness behind two DOI-archived benchmark reports, and why a study premise can be killed in one working day

A benchmark that killed its own hypothesis: strong models re-derive what they can check, and a weak model trusts a stale note over the evidence in front of it

A reproducible benchmark: 42.7% to 98.7% on knowledge-trap questions, and a platform that learns from empty

An Apache-2.0 MCP server that gives frontier models access and grounded context to everything you built

Metadata, lineage, and a business glossary over the whole platform

An insight an analyst shares mid-session becomes a reviewed, signed-off change to the catalog rather than vanishing when the conversation ends.

Steering agents from the server middleware instead of the tool descriptions.

An MCP server faces an agent on both sides. Go lets me leash the agent that writes it and harden the contract the agent reads.

Cloud-grade data platforms from FOSS on Kubernetes, now that agents run the hard part

One server that uses every primitive, runs over both transports, sits behind auth, and is tested end to end, with one annotated trace through all of it

The experimental 2025-11-25 primitive that turns a request into call-now-fetch-later, hand-rolled on the wire because the SDK has not caught up

A server's words flow straight into the model's context. The anatomy of how that becomes an attack, and where the host has to stand to stop it

How a remote server proves who is calling: the discovery chain, resource indicators that bind a token to one server, and the passthrough ban that stops a confused deputy

Structured server logs filtered by severity, and the list_changed notifications that keep a client's view of a server from going stale

The plumbing every MCP primitive shares: how a long call reports progress, how either side cancels one, how a connection is checked, how a list is paged

The server asks a person for input mid-task, with a flat form for ordinary data and a URL handoff for the secrets a form must never touch

The client hands the server a list of directories it is allowed to use, a boundary the server asks for and the client defines and enforces

The protocol inverts. A server asks the client to run the host's model, with no API key of its own and a human able to deny every request

One method that suggests values as a user fills in an argument, and the context field that scopes those suggestions to what is already filled

The third server primitive and the one people misread: templates the user selects, rendered into messages that can embed a server's own resources

The primitive for data the model reads, the counterpart to tools, with URI templates for whole families and live notifications when something changes

Every shape a tool result can take: the five content types, structured output validated against a schema, and the difference between a failed tool and a failed protocol

What the model actually receives when it lists a server's tools, how the schema is built from a Go type, and why a tool's annotations cannot be trusted

A single HTTP path carries the protocol. In 2025-11-25 a session id is optional. In 2026-07-28 it is gone.

The simplest MCP transport, built by hand on both sides with no SDK, plus the stderr rule that silently breaks servers

Vector Search Without the Vector Database, Part 1