ctxt.dev — Context Engineering (EN)

[Читать по-русски](/posts/ru/taskroute-delegation-overhead/) I use Astra in Codex for work beyond software development.

A task is a useful handle.

AI-agent workflows have a boring but important problem: a good process often lives inside one runtime.

A month ago I wrote about [spec drift](/posts/en/spec-drift-living-intent) -- what happens when an agent reads a stale `project.intent.md` and treats wrong facts as ground truth.

Most agent tooling mistakes coordination for reliability.

Last week, our pipeline produced a proofpack with `decision: HUMAN_REVIEW`.

I maintain 12 Claude Code plugins.

Last week I asked Claude to add JWT authentication to a project.

The morning digest surfaced [Leanstral](https://mistral.ai/news/leanstral) -- Mistral's open-source agent for formal verification in Lean 4.

I use three AI coding CLIs daily: Claude Code, Codex CLI, and Gemini CLI.

The first version of Signum ran in a single pass: CONTRACT → EXECUTE → AUDIT → PACK.

You install 16 skills.

We talk a lot about prompts, tools, and evals.

Herald was a Python plugin that collected RSS feeds and Hacker News, clustered articles by title similarity, and generated Markdown briefs.

You asked an AI research assistant a detailed question and got a confident multi-page answer with citations.

AI made writing code mass-affordable.

AI generated a function in seconds.

I wanted my AI agent to know what's happening in tech - without cloud APIs, paid tiers, or data leaving my machine.

You can review an AI diff with three models and still have zero ground truth.

## Problem AI agent tools are scattered.

Everyone who works with AI agents writes instructions for them.

## Problem Git was built for humans.

## Problem You spin up a team of three agents.

## Problem AI agents are the new readers of the web.

Gas Town is an orchestrator by Steve Yegge for parallel management of dozens of Claude Code agents.

## Problem Brainstorming startup ideas is chaotic.

**Disclaimer**: Loadout is my experiment, not a production-ready solution.

## Problem Coding agents were toys for a long time.

## Problem Text-to-SQL seemed like an elegant solution.

## Problem Prompts stopped working reliably.