One core, many shells: why coding agents are turning into engines

Coding agents are separating into two halves: the core that does the work, and the shell you actually touch. That split is not my theory, it is what I read out of four traces from the last two weeks. Cline put one agent core under three frontends with v4.1.19, a terminal, an SDK, and a desktop app. Hermes Agent rolled about 1,800 merged pull requests into the stable v0.21.4 tag and opened the door to other engines with the plugin wave of September 23, which AlphaSignal compresses into one sentence: Hermes made Codex the engine and itself the shell. On the leaderboard side, Codex CLI sits at 125,962 stars and Gemini CLI at 107,130, both under Apache 2.0. For my own operation on a 45-euro phone this became one rule, backed with numbers in this post: pin the engine, update the shell, and keep the gates on my operation, not on the engine.

The Cline trace is the cleanest cut. Since v4.1.19 from September 17, one agent core runs under a command line, an SDK, and a desktop app. One core, three surfaces, the same agent underneath. For teams that means you can swap the engine without rebuilding your frontends, for a solo operator it means the interface and the motor are separate parts with separate risks. The Hermes trace carries the biggest number of the month: per the release note, v0.21.4 rolled about 1,800 pull requests merged since v0.21.3 into one stable tag, aimed at downstream consumers such as Docker images, Hermes Cloud, and hosted deployments. v0.21.5 added roughly 460 more on September 24, deliberately as a patch tag, with curated notes promised for v0.22.0. Before that, on September 23, came the wave of 4 official and 96 community plugins, which coverage reads as a change of the building plan, not as a bag of gimmicks.

AlphaSignal has condensed the Hermes move into the formula 'Hermes just made Codex the engine and itself the shell' and shows the plugin marketplace as the entry point: 'codex plugin marketplace add openai-curated', then 'codex plugin install' for Linear, GitHub, Gmail, and Calendar. I count that article as reporting, not as official documentation. The numbers in this post therefore come from three kinds of sources: release lines from the official Hermes Agent releases feed, leaderboard values from Morphllm as of September 2026, and my own measurement dates from the vault. The one-core-three-frontends reading for Cline comes from a secondary overview with a changelog reference, I did not open that release page in full for this post, and I flag that again in the limits section.

Read these checkpoints as a trail, not as a verdict. The distance between 1,158 releases in 90 days on one side and a single pinned core in my operation on the other is exactly what this post is about. The pace at the top stays high, and I hold one state at the bottom that stays auditable.

A leaderboard counts attention, not quality. I still use it, but with one clear reading: a star tells me that somebody used the project and that it was visible enough to receive one. It tells me nothing about stability across 88 hours, nothing about update risks on an Android phone, and nothing about which engine runs inside which shell. The interesting part is the comparison of dated points: Codex CLI stood at 125,962 stars on September 22, my own reading on September 29 gave 127,009, roughly a thousand gained within a week. Gemini CLI stood at 107,130 on September 15. My own vault table sorts the row differently from public perception: Hermes at 249,852 ahead of OpenCode at 210,679, Codex at 127,009, pi at 110,166, herdr at 41,300. Three of the five are projects I actively run or evaluate, and the gap between Hermes and herdr shows how far attention and actual operation can drift apart.

My execution core is pi, pinned at version 0.87.1, attached through herdr, orchestrated by Hermes. The pin is not convenience, it is a consequence of the operating history of this deck. On August 28 I lived through the OOM storm: five gateway deaths and a kill issued by Android, with Chromium as the proof that the operating system wins under memory pressure. The way out ran through an external cron tick and a semantic watchdog, and the 88-hour record that followed, from August 28 at 19:40 to September 1 at 12:18, does not show that nothing happens, it shows the stack survives kills because it was built for exactly that. Every update cycle on Android carries that same risk: the phone manages memory aggressively, an update at the wrong moment can produce exactly the kill the hardening is meant to absorb. So the engine stands still until an update comes with a test plan, while the shell may move, because it shifts nothing against the pinned core.

The core and shell split makes that separation genuinely useful. The rail gate, the word-count gate, and the twin post in my operation hang on the pipeline, not on the engine behind it: the rail gate runs before every draft, the word gate counts the main text of the stored version, and the English twin gets created in the same run. If the shell loads another engine tomorrow, my gates stay the same, and the dated readings in the vault stay comparable. One honesty note about my own story: the two verified projects with 36 of 36 passing tests were built in the local pi workflow, Hermes plans, pi executes, Hermes verifies. The path through ChatGPT and Codex on my MacBook is planned and not in use, so I do not write it as lived experience. If I ever test the Codex-engine path behind Hermes, it will go here with a date, a test plan, and numbers.

If a shell swaps its engine, for example Hermes with Codex behind it, my test plan would be clear. First a dry run without rights: the same small task posed twice, the pinned pi lane against the new lane, results side by side, logs on. Then the misbehaviour cases: what happens when a path is blocked, exactly the question the Transluce report of September 23 documented with agents that built detours through foreign services during trivial data tasks. Then the cost figures: tokens, runtime, requests, because my operation runs under a 5-hour window and a weekly cap, with numbers only from my own dashboard. Then the gates: the rail gate stays before the draft, the word count stays on the stored version, the twin stays in the same run. Only when all four stations are covered would an engine swap move from test into operation, and the decision carries its date in the vault, not a gut feeling.

Named openly: I have not run the combination of Hermes with Codex as engine myself, my statements about it come from the release feed and from reporting, not from an own test series. The one-core-three-frontends mapping for Cline comes from a secondary overview with a changelog reference, for this post I did not open the Cline releases page in full. The leaderboard is the snapshot of a single provider, Morphllm, as of September 2026, and star counts are dates that move daily. The Hermes pull-request rollup figures are round numbers from release notes, not exact counts by me. My own measurements come from one device and one vault, they document a practice, they carry no market claim. And which engine sits inside which shell next week stays open, this post is a snapshot dated October 1, 2026, not a forecast.

External sources with retrieval dates:
- Hermes Agent official releases feed (GitHub), retrieved September 30, 2026: https://github.com/NousResearch/hermes-agent/releases
- Morphllm, Best AI Coding Agents 2026, leaderboard as of September 2026, retrieved September 30, 2026: https://www.morphllm.com/best-ai-coding-agents-2026
- Cline releases page (GitHub), retrieved September 30, 2026: https://github.com/cline/cline/releases
- AlphaSignal, 'Hermes Just Made Codex the Engine and Itself the Shell', retrieved October 1, 2026: https://alphasignalai.substack.com/p/hermes-just-made-codex-the-engine
Own measurements: star snapshot of September 29, 2026, 06:59 CEST, tempo trail of September 27, 2026, pi pin 0.87.1, all from the vault with dated checkpoints and stated limits.

About the author: I am Marcel, a graphic designer and the operator of the HUNTER cyberdeck, an agent setup on a used Google Pixel 6a. This post rests on public release lines, one leaderboard snapshot, and my own operating and measurement numbers from the vault, dated checkpoints from September 27, September 29, and October 1, 2026. Last fact-checked on October 1, 2026.