11 entries
A local guardrail can approve a task that the server will cut off before its warning threshold. Reconcile server windows, per-model limits and locally valued usage before trusting it. The September observations show how a stale catalog can break that protection.
Keep conventions, acceptance criteria and tests in the repository to make your agent replaceable. Hooks remain adapters that need porting and testing; an unused fallback route has an upkeep cost too.
Define the reference, quality criteria and time categories before starting your AI pilot. Compare comparable tasks and include both correction and review. The protocol below remains a proposed method, with no controlled comparison results or demonstrated business ROI.
A team can start with four rules: limit what the agent can touch, version shared conventions, automatically block verifiable mistakes, and keep human approval before delivery. Roll the framework out on one pilot project, then expand it only when the measurements show that it actually helps.
herdr is a background server that owns your coding agents' terminals: every pane exposes a state (idle, working, blocked, done), and the CLI and socket let the agent itself drive the herd. This guide says what I do with it daily, the split/wait/read flow, and the verified traps.
omp (Oh My Pi) is more than a coding-agent CLI: it's a harness where skills, hooks, subagents and per-model budget live in the config. Used daily, this guide shows the real setup, the gates, and the first steps to migrate without rewriting everything.
Everyone builds agents. Almost nobody writes the tests that decide whether they can ship. An agent isn't judged by feel: you put it under gate, like code. Here's what that looks like in practice, indirect injection included, the test everyone is missing.
A local proxy (RTK) shows 4.4M "saved" tokens (45.3%) across 9,892 cumulative commands (global scope). That figure reads like a lower bill; that reading is incomplete. That "saving" only exists because a harness (hooks, gates, routing) produces it and keeps producing it. A deliberate share of my AI budget has to fund that upkeep, not just more generation tokens.
A counter built into the tool can't measure the tool's savings: it sees neither the counterfactual, nor the cache, nor retries. Measured against my invoice, the real gain of four compression tools lands between 0 and 15%, while cache reads, meanwhile, make up 97% of tokens and re-read every token ~36 times.
The agent types code faster than I do. That was never the bottleneck. Across my tracked sessions, most last under ten minutes, half go down the wrong approach, and a small minority end in a commit. The blocker didn't disappear — it moved, from writing to framing.
A rule written in a CLAUDE.md, the agent eventually forgets: a few tens of thousands of tokens later, it's no longer in its window. A guardrail — a hook that refuses the action — doesn't forget. I moved the rules that matter from markdown to about fifteen hooks. Here's what it changes, and what it really costs.