Notes
Working notes on agentic engineering, AI, and quantitative finance. Less polished than a paper, more honest than a tweet.
-
decision-models
How to Use Jev in 2026
Jev returns typed probabilities instead of prose. A practical guide to the three primitives, confidence gating, real costs, and where it breaks.
-
model-selection
Claude Sonnet 5.5 vs Opus 5.5: Cheaper Per Task
Sonnet 5.5 costs half of Opus 5.5 per token, yet the bill per finished task can land in the same place. The mechanism, and when Sonnet genuinely wins.
-
business
Claude for Business: SEO, Outreach, Where It Breaks
Anthropic just shipped Claude Cowork for non-coding knowledge work. Here is what it reliably does for sales and marketing, and the data problems that stop it.
-
coding-agents
What Are AI Coding Agents? A Practical Guide
Learn how AI coding agents work, when to use them, and how to build one with tools, repository context, tests, and bounded execution.
-
frontier-models
Claude Opus 5.5 Cooked While the Frontier Kept Shipping
Claude Opus 5.5 leads an independent intelligence index and cuts token prices, even as its launch lands inside a fast-moving frontier price war.
-
decision-models
Jev: Where a Fast Decision Model Fits in an Agent Stack
Jev swaps open-ended generation for typed probability outputs. A guide to agent workflows, efficiency claims, local models, and the limits of the evidence.
-
cost-economics
Cheap Intelligence Is a Cost-Curve Problem
Cheap AI is commoditizing capable work, shifting competition toward cost per completed task and exposing incentives behind calls to slow frontier development.
-
memory
Agent Memory That Beats Grep: How I Built Heimdall
Why agents redo work, how I built Heimdall — a verified, self-healing knowledge layer for AI coding agents — and how it compares to graphify and Graft.
-
foundation-models
DeepSeek V4 Pro 0813: Frontier Coding at Commodity Price
DeepSeek's GA release closes the agentic coding gap with Claude and GPT — Terminal-Bench 2.1 at 87.9, DeepSWE up 5x — for roughly 1/12th the price.
-
token-economics
DeepSeek V4 Flash 0731: Loop Economics and the Price Cascade
V4 Flash 0731 sits one point behind GPT-5.6 Luna at a fraction of the cost. Loop math, the best harnesses, and the price cascade is a gift to indie developers.
-
hooks
Deterministic Tool Rewriting: Hooks vs Model Discipline
Deterministic tool rewriting moves safety rules out of the prompt and into code. Why hooks beat model discipline for enforcing agent behavior, with examples.
-
foundation-models
DeepSeek V4 Flash 0731 vs GPT-5.6 Luna: Scores, Prices
DeepSeek's official V4 Flash 0731 beta and OpenAI's 80% Luna price cut redraw the practical price-performance frontier for production agent fleets.
-
routing
Routing Through Free-Tier LLMs Without Hitting Limits
Free tier LLM routing breaks when one provider rate limits. How to spread agent traffic across free providers, detect throttling, and fail over cleanly.
-
code-review
AI Code Review: Alibaba's LLM and Deterministic Checks
AI code review that pairs Alibaba's deterministic-pipeline checks with an LLM agent for line-level review: what it catches, and where it falls short.
-
code-review
Making LLM Code Review Actually Catch Things
LLM code review only catches real bugs when it has tests and diff context to work from. A practical setup for code review that finds issues worth reading.
-
foundation-models
GPT-5.6 Luna Is the Value Tier. Terra Is Not Useless.
GPT-5.6 Luna is the rational default for bounded agent work, while Terra remains the escalation tier for ambiguity, mixed context, and harder coding loops.
-
open-weights
Open Source Frontier: Giant Models, Lower-Cost Coding
Kimi K3, Qwen3.8-Max-Preview, GLM-5.2, OpenCode Go, Qoder, and ClinePass show the frontier shifting from model announcements to usable agent economics.
-
news-roundup
Model Release Roundup: What Actually Changed
Model release roundup: which launches actually moved capability and price this cycle, which were repackaging, and what changed for teams picking a model.
-
tool-use
Tool-Use Loops: The Failure Modes Nobody Documents
Tool use loops fail predictably: retried calls, malformed output, lost state, runaway recursion. The failure modes, how to spot each, and how to stop them.
-
prompt-caching
What "Prompt Cache Warm" Means (and How to Keep It Warm)
"Prompt cache warm" means your cached prefix was reused — no re-processing, lower cost and latency. What it means, why it expires, and how to keep it warm.
-
subagent
Fan Out Subagents: Patterns That Ship (and Break)
"Fan out subagents" means splitting one task across parallel workers. When fanout beats pipelines, the 10-agent scaling wall, and how to combine both.
-
news-roundup
Trending AI Repos Worth Cloning This Week
Trending AI repos worth cloning this week: what each project actually does, how active it is, and whether it solves a problem your agent stack already has.
-
claude-code
This Week in Claude Code: Features Worth Trying
This week in Claude Code: what changed in the CLI, auto mode and agent behavior, the features worth turning on, and which are too rough to rely on.
-
coding-agents
AI Coding Agents: Claude Code vs Aider, Cursor, and Codex
Compare leading AI coding agents across autonomy, repo context, editing, test execution, and cost. See which harness fits your development workflow.
-
context-window
Context Window Management for Real Agent Sessions
Context window management decides which tokens survive a long agent session. Tactics for compaction, retrieval, and summarization that hold up in real runs.
-
mcp
MCP Server Design Patterns That Actually Hold Up
MCP server design decides whether an agent can use your tools reliably. Patterns for tool naming, output shape, error reporting, and context cost that hold up.
-
hooks
Hooks vs Skills vs Subagents: Choose the Right Primitive
Hooks run at fixed points, skills load on demand, subagents run in their own context. Which Claude Code primitive fits the job, with clear selection rules.
-
token-economics
Token Economics of Long-Running Agent Loops
Token economics of agent loops: where long-running agents burn budget, why loops cost more than single calls, and the levers that bring cost per task down.
-
claude-code
How Claude Code's Skills System Actually Works
Claude Code skills package reusable instructions agents load on demand. How the skills system works, where it beats hooks and subagents, and when to use each.
-
agentic-engineering
Agentic AI: Production Patterns That Actually Hold Up
Learn the agentic AI engineering patterns that survive production: context budgets, tool design, planner-executor splits, evals, and failure recovery.
-
frontier-models
Frontier AI in 2026, what actually changed and what did not
A working note on the shifts in the AI frontier through mid 2026. Long context, agentic capability, open weights, and the parts the headlines got wrong.
-
finance
LLM Agents in Quantitative Finance: Where They Pay Off
A working note on where coding agents earn their keep in equity research, forensic accounting, and DCF construction, and where they still do not.