# Agent-world user feedback → product decisions

Observed October 21, 2026. This is a small, deliberately targeted qualitative sample, not a representative survey or a claim of competitor-wide failure. We read the actual review tab and maintainer response, plus firsthand issue reports. We did not reproduce competitor bugs. Issue status does not establish whether the latest release still has a problem. No reviewer was contacted.

## Preserve what people enjoy

The [Pixel Agents VS Code review tab](https://marketplace.visualstudio.com/items?itemName=pablodelucca.pixel-agents&ssr=false#review-details) displayed five reviews: four positive and one critical, with an aggregate 4.2/5 on the page. Positive comments valued enjoyment and awareness of agent activity. One favorable reviewer wanted more visible status and reported missing agents in a multi-role, single-terminal workflow. Enthusiasm is an engagement signal, not evidence of recurring paid demand.

A separate [Stormies JetBrains listing](https://plugins.jetbrains.com/plugin/30474-stormies--pixel-agents) contains a positive comment about watching multiple Claude sessions and a complaint involving dot-prefixed project paths. Stormies is a different implementation; do not combine its ratings with Pixel Agents.

Implications: retain the playful world, make state understandable without hovering, represent logical agents explicitly rather than assuming one terminal equals one agent, and avoid coupling identity to project-path parsing. A roster and focused inspector should remain useful when the world is crowded.

## Treat trust failures as product failures

The critical VS Code review reported overwritten agent settings. The maintainer acknowledged the failure and stated it was fixed in **v1.4.1**, adding consent, merging, backup and safer uninstall behavior. This is a historical incident with a reported fix, not a claim that today's version still overwrites settings. Forge's connector takes another route: no hooks or agent-settings edits; an explicitly selected local module and its own owner-only credential file.

## Reported pain → release requirements

| Firsthand source | Reported experience | Forge decision and acceptance check |
|---|---|---|
| [Pixel Agents #413](https://github.com/pixel-agents-hq/pixel-agents/issues/413) | Stale sessions and unfinished activity labels persisted | Explicit runtime heartbeat, offline state, fenced run lease and interrupted outcome. Stop a connector during work: show uncertainty, reject the old lease and never automatically rerun paid work. |
| [Pixel Agents #251](https://github.com/pixel-agents-hq/pixel-agents/issues/251) | Character click selected the wrong session after another session was cleared | Bind actions to workspace, run, registered handler and lease IDs. Test cross-world credentials and stale-run actions. Never resolve by folder or display name. |
| [Pixel Agents #272](https://github.com/pixel-agents-hq/pixel-agents/issues/272) | Crowded characters produced overlapping labels | Keep a keyboard-accessible roster and inspector; cap world capacity and test a 64-agent scene. Do not claim a fully solved spatial layout before that visual test. |
| [Pixel Agents #286](https://github.com/pixel-agents-hq/pixel-agents/issues/286) | Request for a more noticeable permission signal | Surface failures and pending acknowledgements in a persistent attention area; add optional alerts only for actionable events. Actual approval routing is a separate feature, not an animated implication. |
| [Star Office UI #108](https://github.com/ringhyacinth/Star-Office-UI/issues/108) | Guest detail/presence did not match synced data | Pair a runtime, verify IDs, publish negotiated capabilities, and show its connection separately from a task's reported completion. Test API-to-render mapping. |
| [Pixel Agents #220](https://github.com/pixel-agents-hq/pixel-agents/issues/220) | Closed terminals left orphan characters | Preserve history while distinguishing offline from working. A missing process must not look live because its last animation continues. |

The common qualitative theme across these selected reports is mismatch between an attractive representation and its underlying runtime. It supports prioritizing trustworthy identity and control, but does not establish how frequent these issues are or whether users would switch to Forge.

## Differentiation worth earning

Pair existing runtimes without editing their settings; launch only user-registered handlers; show requested versus acknowledged controls; retain run-specific results; produce conservative evidence briefs; and eventually support selective client-facing handoffs. Several competitors already offer parts of this vision. Our claim must become a demonstrably better combined workflow for a specific customer, rather than a claim that no one has thought of it.

The next user study should test five tasks with a connected team: connect it, launch a task, identify the active agent, pause/resume at a boundary, and recover understanding after disconnection. Record success, errors and time. Compare against an ordinary task list and a relevant alternative. Ask about actual recent use and spending before proposing prices. Do not substitute positive comments, installs or stars for renewal evidence.

## Sample and limitations

Primary review page: five visible VS Code reviews, including the response to the critical review. Adjacent implementation: two substantive Stormies comments used qualitatively. Issue sample: the six linked reports selected for identity, presence, visibility and control relevance. No exhaustive scraping, sentiment percentage for the category, randomized sampling, bug reproduction, or willingness-to-pay conclusion is claimed. The research specifically excludes unrelated products sharing the Pixel Agents name and automated review-score sites.