Jcode and Ponytail: Tiny Coding Agents With Outsized Habits

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

Jcode arrived in the notes as a Rust-rewritten coding-agent harness — claimed roughly twenty times lighter than Claude Code and sixty times faster than Codex — running ten parallel specialist agents. Ponytail arrived as the YAGNI plugin that makes agents ask whether code should exist at all. I do not need every benchmark to be gospel. I need the habits: lighter loops, parallel specialists with a merge plan, and a cultural bias against inventing abstractions nobody requested.

Lightweight harness, heavyweight manners

If Jcode's performance claims hold on my machine, it becomes a spike tool for parallel exploration. If they do not, the manners still transfer: keep the harness thin, keep specialists narrow, keep the orchestrator boring. A heavyweight UI that hides the plan is slower in human time even when the tokens fly.

I benchmark on my repos, not on vendor demos. My repos have the ugly fixtures that demos omit.

Ten specialists need a merge protocol

Front-end, back-end, tests, research — parallel agents are only free if their outputs reconcile. I require a shared schema, a single integrator role, and a rule that specialists cannot merge to main. Parallel without protocol is a race condition with a marketing name.

Time boxes per specialist beat open-ended "keep improving" loops that never converge.

Ponytail's question as default culture

Ask whether the code needs to exist. Average code reductions around fifty-four percent in the notes — sometimes far more — with claimed cost and speed wins while keeping security checks. Even without the plugin, I paste the question into AGENTS.md.

Agents love building frameworks for one-off problems. Ponytail energy is how I keep the diff reviewable and the junior teammates able to learn from what landed.

When tiny agents beat famous ones

Tiny harnesses win when the task is local, the tests are fast, and the blast radius is a branch. Famous cloud agents win when you need connectors, long memory, or organisational SSO. I pick by blast radius, not by Twitter heat.

Clients hear this as: we will not pay frontier prices to rename a button. We will pay them to redesign auth.

Security checks stay non-optional

YAGNI is not an excuse to skip threat review. Ponytail-style reductions that drop security tests fail my gate even if the demo looks sleek. Less code with fewer checks is not a win; it is a smaller attack surface with less visibility — sometimes worse.

I keep a minimum security skill that may not be YAGNI'd away: secret scanning, dependency audit, auth diff review.

Teaching the team the habit

Workshops: run the same prompt with and without an explicit YAGNI instruction, compare diff sizes, compare review time. People believe measurements more than sermons. Then we add the instruction to the shared skills pack.

Stars in nine days for restraint tools tell me the market is hungry for less. I sell less, carefully.

Pilot shape

One lightweight harness on a non-critical repo, three specialists max, Ponytail-or-equivalent question required, human integrator, two-week churn metric. Expand only if review time drops without defect rate rising.

If defect rate rises, we kill the parallel party and go back to one careful agent. Ego does not get a vote.

What I keep saying

Fast is optional. Necessary is mandatory. Parallel is optional. Merge protocol is mandatory. Tiny can be mighty when it refuses to invent work. That is the whole post, and it fits on a sticky note above my keyboard.

Churn budget

Each specialist agent gets a maximum lines-touched budget relative to the ticket. Blow the budget and the run stops for human review. Jcode speed does not get a carve-out. Ponytail questions trigger before any budget increase.

Budgets sound corporate. They feel like calm when ten specialists want to "improve" everything they can see.

Security skill that cannot be YAGNI'd

Listed by name in AGENTS.md: secret scan, dependency audit, auth diff review. Ponytail may challenge other code. It may not challenge those three. Tiny harnesses still carry adult responsibilities.

Specialist merge theatre ban

Specialists never demo directly to stakeholders. The integrator narrates the merged result. Jcode parallel speed stays behind the curtain so Ponytail questions and security skills remain visible to humans.