Version 2 feedback

[matt17r]

While I like the new "roles" in Agent OS 2, man it chews up my tokens!!! Agent OS has always made me hit the 5 hour limits on my Pro tier after 3 or 4 hours of back and forth but now in version 2 I can hit my 5 hour limit in well under 2 hours!

I'm also finding the spec commands a little confusing... I just came back to Claude after a while expecting it to have created the next spec, only to see it spent 10 minutes going through an already completed spec and validating it for the third time. Apparently "create spec" doesn't mean "create spec". It later told me I should have used "init spec" but the auto completion name is "new spec". "Create", "new" and "init" all seem pretty synonymous to me and not at all intuitive.

Then when I did tell it to create new? a new spec, it asked "What feature would you like to create[sic] a spec for?". Felt a little silly that it didn't just start working on the next feature in the roadmap 🤷‍♂️

[matt17r]

Upon further reflection, I think you've already come up with better names for the roles... could you not use those for the spec commands?

• /agent-os:research-spec
• /agent-os:write-spec

[mishakalyna]

In version 1.1.0, the following improvements were noted:

Improved
• Context efficiency by 60–80% through conditional loading and lite file versions
• Reduced duplication when files are referenced multiple times in a workflow
• Clearer separation between task-specific and full test suite execution
• More intelligent file loading that checks the current context before reading
• Better organization of code style rules with language-specific files

I think these optimizations might have influenced the decline in context efficiency starting from version 2.0. My assumptions are:
• The absence of a context-fetcher sub-agent due to framework architecture changes
• Possible multi-agent parallelism, which produces faster and higher-quality results — but also consumes tokens faster
• Additional verifiers that ensure higher quality after implementation, yet increase token consumption
• New UI-related capabilities (screenshot validation, MCP Playwright Browser integration), which also demand more tokens
• More extensive documentation
• No lite documentation versions
• Excessive number of roles — each needs to read documentation before starting, consuming a large portion of context. Reducing roles could save tokens but may harm quality.

We should think about how to minimize token usage. The only levers we have are reducing documentation and role count — but doing so would essentially bring us back to the earlier, simpler concept that arguably worked better.

Regarding the commands “new-spec” and “create-spec”, they indeed cause confusion. It might be better to simplify the flow to: create-spec → create-tasks → implement

Also, in single-agent mode, there are now too many commands, which complicates the workflow.
Since many IDEs and CLIs allow command invocation via slash commands (”/”), having fewer and simpler commands here would significantly improve usability.