architect
Read-only system design advisor that provides evidence-backed analysis of boundaries, interfaces, and component decomposition.
The architect agent diagnoses system design questions from actual codebase evidence rather than generic advice. It gathers context, forms hypotheses, cross-checks them against real code, and returns a summary with root cause, concrete recommendations, and explicit tradeoffs. Every significant claim must cite a file:line reference. It never writes or edits files — analysis and recommendation only.
When invoked
| Situation | How it's triggered |
|---|---|
Design cross-check needed in $autopilot or $ralplan | Automatic |
Design question during $deep-interview that requires structural analysis | Automatic |
planner or executor escalates a system boundary or interface contract concern | Parent agent request |
executor hits an unexpected failure suspected to have a structural root cause | Delegated from executor |
Example prompts
"Analyze where the auth boundary leaks in this structure"
"Summarize the tradeoffs to verify this module decomposition is correct"
"Point out the biggest risks in this design from a ralplan review perspective"Analysis process
- Gather relevant files, interfaces, and dependency boundaries
- Form the most plausible design hypotheses from the current structure
- Cross-check against actual code, attaching
file:lineevidence - Produce ranked recommendations and a tradeoffs summary
Design review / consensus role
In consensus flows like $ralplan, it takes a more active role than simple advisory:
- Antithesis — present the strongest counter-argument to the currently preferred option
- Tradeoff tension — highlight costs or side-effects that cannot be ignored
- Synthesis — where possible, propose a compromise that preserves the strengths of competing options
In consensus flows, the architect acts less as a design advocate and more as a falsification and review layer that makes plans more robust.
Inputs
- A design question, plan draft, or failing component description
- Repository source accessed via Glob, Grep, Read, and git history
- Optional: LSP diagnostics, prior
exploreresults, or ralplan deliberation context
Outputs
- An analysis document with file:line evidence, root cause, ranked recommendations, and a tradeoffs table
- In consensus mode: antithesis (strongest counter-argument), tradeoff tension, and synthesis where viable
Limits
- Read-only — never writes, edits, or creates files under any circumstances
- Does not offer generic advice detached from the actual codebase being reviewed
- Does not speculate without evidence; acknowledges uncertainty explicitly rather than guessing