R&D Club / Journal12 / Published notes

Notes from the work, not content about the work.

Research positions, engineering decisions, failed assumptions, and the systems that survived contact with production.

Published notes

01
Agent security4 min read

Agent Sandboxes Have Real Attack Graphs

Blocking direct Internet access does not isolate an agent when package proxies, shared caches, credentials, and paths into adjacent infrastructure remain reachable.

Read note
02
Speech systems4 min read

An ASR Transcript Should Be Revisable State

A streaming transcript is not a finished document. Later speech can change earlier meaning, while unconditional cleanup can delete the condition that matters.

Read note
03
GUI agents4 min read

A Screenshot Does Not Prove GUI-Agent Success

The screen can look correct while a file is unsaved, a setting never applied, or a visual judge mistook an intermediate frame for the final state.

Read note
04
Agent economics4 min read

Model Routing Needs Cost per Successful Task

Token price excludes retries, cache behavior, tool use, and failures. A router should select models by measured outcomes on the work it actually performs.

Read note
05
Multi-agent systems4 min read

More Agents Do Not Create Independence

Similar agents can copy one error, lose evidence during handoffs, and mistake agreement for confirmation. Reliability depends on independent checks, not headcount.

Read note
06
Production agents5 min read

A Prompt Is Not a Policy Engine

Putting a long operating manual into the system prompt does not keep an agent inside the allowed workflow. Checks, transitions, and prohibitions need executable enforcement.

Read note
07
Automated R&D4 min read

Self-Improving Agents Need a Best-So-Far

The latest training run, dataset, or skill revision is often worse than an earlier one. Without holdouts, stopping rules, and rollback, improvement becomes regression.

Read note
08
Agent security6 min read

SKILL.md Is an Executable Supply-Chain Dependency

A coding agent reads a skill as instruction and executes it with its own authority. Code review and static scanning see only half of that attack surface.

Read note
09
Agent training4 min read

Synthetic Environments Can Train the Wrong World

Scaling artificial tasks helps only when the environment preserves causal logic, the verifier reads actual state, and defects in the environment are repaired.

Read note
10
Agent architecture4 min read

The Harness Is Part of the Model

The same model can produce different results when context assembly, tool order, retries, and verification change. Comparing model names alone misses the system that executes the task.

Read note
11
Voice agents4 min read

A Voice Agent Must Know When Not to Speak

Detecting speech is not enough. An agent must distinguish an utterance addressed to it from side conversation, an unfinished thought, and a revision of prior intent.

Read note
12
Research practice4 min read

Most AI R&D Is Just Expensive API Integration

Integration can be useful. Research begins when the implementation path is not obvious and the evidence can change the direction.

Read note