Notes from the floor
Blog
Product updates and engineering notes on coding agents, opencode subagents, and running a virtual office in the terminal.
AGENTS.md best practices: the agent reads it before your code
AGENTS.md, CLAUDE.md, and their cousins decide what your coding agent knows before it touches a file — the landscape, the numbers, and the practices that keep the repo's second README from lying.
The code got cheap this year. The supervision didn't.
Coding agents went from autocomplete to async workers you hand tickets to, and the bottleneck moved from model quality to supervision — what actually changed in 2026, and how to stay oriented when several run at once.
How to read coding agent benchmarks without getting played
Every lab's chart says its own model is the best coding agent. What SWE-bench, Terminal-Bench, and the arenas actually measure, why the same model swings twenty points between harnesses, and how to pick with a test you run yourself.
A permission is not a question
Shell allow, product decision, and status are three interrupts. One popup trains you to rubber-stamp a breaking change like it was git status. Here is how the mix-up happens, and how to unmix it.
Running more than one coding agent at a time
A second terminal feels like speed. Then you have three sessions, two waiting on you, and no idea who owns auth. What actually breaks — and the few habits that keep the afternoon intact.
Start with a workflow. Agents come later.
Most 'multi-agent' coding setups are a scripted path with extra avatars. Here is the ladder we use — and the afternoon that taught us to stop at step one.
AGENTS.md best practices: the agent reads it before your code
AGENTS.md, CLAUDE.md, and their cousins decide what your coding agent knows before it touches a file — the landscape, the numbers, and the practices that keep the repo's second README from lying.
The code got cheap this year. The supervision didn't.
Coding agents went from autocomplete to async workers you hand tickets to, and the bottleneck moved from model quality to supervision — what actually changed in 2026, and how to stay oriented when several run at once.
Claude Code vs Codex vs opencode vs Antigravity is the wrong fight
Everyone argues about the best AI coding CLI of 2026; the afternoons that actually ship give Claude Code, Codex, opencode, and Google Antigravity different jobs and never let them share a checkout.
Grok coding is real now. Here's where it fits and where it doesn't.
xAI ships a new Grok every few weeks and the names stopped helping months ago — here's the current coding lineup, what it costs, what it's genuinely good at, and the jobs to give to someone else.
The office doesn't care which brain the boss has
Claude Code is now a supported LLM transport next to opencode. Same floor, same roster, same board, same ledger — you pick which CLI the boss thinks with at boot.
What Kimi K3 actually changes when you run coding agents all day
Moonshot's Kimi K3 puts a 2.8T open-weights coding model with a 1M-token window behind an Anthropic-compatible endpoint, so swapping it into your terminal is an env var — the real questions are the bill, the thinking tax, and which jobs you stop giving your most expensive model.
How to read coding agent benchmarks without getting played
Every lab's chart says its own model is the best coding agent. What SWE-bench, Terminal-Bench, and the arenas actually measure, why the same model swings twenty points between harnesses, and how to pick with a test you run yourself.
A permission is not a question
Shell allow, product decision, and status are three interrupts. One popup trains you to rubber-stamp a breaking change like it was git status. Here is how the mix-up happens, and how to unmix it.
Running more than one coding agent at a time
A second terminal feels like speed. Then you have three sessions, two waiting on you, and no idea who owns auth. What actually breaks — and the few habits that keep the afternoon intact.
Start with a workflow. Agents come later.
Most 'multi-agent' coding setups are a scripted path with extra avatars. Here is the ladder we use — and the afternoon that taught us to stop at step one.
You launched the subagents. Now you have to watch them.
Spawning specialists is easy. The job that remains is supervision: diffs you can find, asks that wait instead of freeze, and a handoff you would accept from a coworker.
If the UI can lie, you will believe it
Agent workspaces fail when the picture is prettier than the log. A test: every animation should map to an event you could write to JSONL — and a labeled demo is the only honest fake.
Don't let approvals steal the screen
Four agents, four 'allow this?', you are a clicker. Here is why modal prompts serialize a parallel shift — and the queue habits that give attention back without going full yes-to-all.
A small team's operating system for coding agents
Attention is the scarce resource, not tokens. Four conditions that let a solo developer get more attempts from agents without becoming their full-time supervisor — and when the OS is overhead.
Put the workspace where the agents already are
Coding agents live in the terminal. A dashboard that 'manages' them is another tab to lose. Attach to the session you have — restore yesterday — don't kidnap the runtime.
Memory is whether tomorrow inherits today
A capable agent without a record is a contractor who forgets the meeting. Continuity is inspectable state — notes, boards, session IDs — not a model that 'improves itself' in secret.