# Articles

Analysis of what users write in public, signed and dated, and separate from the benchmark data.

- [Grok Bot users made it a dispatcher for Claude Code and Codex before xAI's routing](https://feedbackbench.com/articles/grok-bot-users-made-it-a-dispatcher-for-claude-code-and-codex-before-xais-routing.md) (Personal agents, Grok Bot, 17 October 2026): In seven weeks of posts, 498 people praised running Grok Bot as a team of bots, each with its own job, and Grok Bot is 1st of 5 on multi-agent teams. Separately, 202 people made Grok Bot a dispatcher for agents from other makers, most often Claude Code and Codex, and 134 of them did so before Musk's 7 October post brought other makers' models into Grok Bot. One in five say they do it to save Grok Bot's allowance or use subscriptions they already pay for.
- [Hermes Agent users choose their models, and Grok Bot users ask to do the same](https://feedbackbench.com/articles/hermes-agent-users-choose-their-models-and-grok-bot-users-ask-to-do-the-same.md) (Personal agents, Hermes Agent, 17 October 2026): In seven weeks of posts, 128 people described choosing, switching or mixing the models under Hermes Agent, which is 1st of 5 on model choice. In the same weeks 57 users of Grok Bot, which has no model picker in its settings, asked to pick, pin or see it, and some named Hermes as the example.
- [Gemini Spark users trust it to file their records; it pauses before deleting](https://feedbackbench.com/articles/gemini-spark-users-trust-it-to-file-their-records-it-pauses-before-deleting.md) (Personal agents, Gemini Spark, 17 October 2026): In the posts we read, 23 people praised Gemini Spark for building up the records they keep, mostly with Google, and 22 of them describe adding, sorting or filing. Where users find it holding back is bulk deletion and some scheduled runs, which stop for a yes one item at a time, and Google already exempts bulk changes in Google Tasks.
- [Users fault personal agents more for holding back than for overstepping](https://feedbackbench.com/articles/users-fault-personal-agents-more-for-holding-back-than-for-overstepping.md) (Personal agents, 17 October 2026): Users of nine personal agents fault them more for asking again and for staying silent than for acting on their own. The benchmarks the field cites mostly find the opposite failure, one test at a time.
- [A personal agent's false “done” surfaces when someone looks, or when it costs](https://feedbackbench.com/articles/a-personal-agents-false-done-surfaces-when-someone-looks-or-when-it-costs.md) (Personal agents, 17 October 2026): We read every post Feedback Bench ties to false reports. Of 63 people who said how they found out, 32 had gone and checked; 22 learned it from an alarm that rang or a report that never came, and 15 of those had handed the agent a scheduled or background task.
- [Zed users praise its speed more often than users of any other coding agent](https://feedbackbench.com/coding-agents/articles/zed-users-praise-its-speed-more-often-than-users-of-any-other-coding-agent.md) (Coding agents, Zed, 15 October 2026): In seven weeks of posts, 81 people praised how fast Zed opens, types and responds, against 16 who complained. Per judging post, that is the highest rate of speed praise among 17 coding agents, and 20 of those users praise it while agents write the code they then read in Zed.
- [OpenCode is 1st of 12 on plan value, as the $10 worker beside a frontier model](https://feedbackbench.com/coding-agents/articles/opencode-is-1st-of-12-on-plan-value-as-the-10-worker-beside-a-frontier-model.md) (Coding agents, OpenCode, 15 October 2026): More than 260 people say from their own use that OpenCode Go buys a lot of work for its price, and OpenCode is 1st of 12 agents on plan value. Many keep a Codex or Claude plan and give the routine coding to Go's open models.
- [Amp's users praise its cloud machines at a higher rate than any rival's](https://feedbackbench.com/coding-agents/articles/amps-users-praise-its-cloud-machines-at-a-higher-rate-than-any-rivals.md) (Coding agents, Amp, 15 October 2026): Of the 1,323 people who posted about Amp in seven weeks, 131 praised orbs from their own use. They describe work that kept running through power cuts and trips, local setups they stopped maintaining, and rival cloud agents that fell short when they tried them.
- [Pi's users ask it to extend and repair itself, and it can read its own manual](https://feedbackbench.com/coding-agents/articles/pis-users-ask-it-to-extend-and-repair-itself-and-it-can-read-its-own-manual.md) (Coding agents, Pi, 15 October 2026): At least 23 people who judged Pi in seven weeks say they had it build, fix or reconfigure part of itself, and 9.8% of everyone judging Pi praised its extensions, the highest share of 17 coding agents. The source shows why it works: the default system prompt points the model at Pi's own docs and example extensions.
- [Cursor is 1st of 5 on letting developers see and approve every change an agent makes](https://feedbackbench.com/coding-agents/articles/cursor-is-1st-of-5-on-letting-developers-see-and-approve-every-change-an-agent-makes.md) (Coding agents, Cursor, 15 October 2026): Cursor is 1st of 5 on Feedback Bench for reviewing and approving the agent's changes, and 2nd of 7 on IDE integration, behind Claude Code. The 28 users who praise it first-hand mostly measure it against rival agents, and some who left say it is the one thing they miss.
- [A coding agent should not be the only judge of its own “done”](https://feedbackbench.com/coding-agents/articles/a-coding-agent-should-not-be-the-only-judge-of-its-own-done.md) (Coding agents, 15 October 2026): On Feedback Bench, 708 of the 738 user-weeks that rate whether an agent claims work is done when it is not are complaints. In users' accounts of checks on a “done”, the same model checking itself usually let a false claim through, while a different model, or a check that read the repository or the screen, usually caught it.