The human seat never empties

Now let me review the actual job postings.I've reviewed the record behind Raza-SOR (what an agent built with the /build-your-agent method actually is) and each posting you linked. Quick notes before the table:

What I evaluated against. Per the governed record, a Raza-SOR worker is a text-only agent on the OpenAI Agents SDK, in Python, wrapped in a FastAPI harness, built into a container image, deployed to Vercel — with typed function tools, sessions, a guardrail, a human approval gate on any tool that changes something, streaming, and tracing, plus the surrounding discipline: delegation map, spec, governance record, and an eval suite that grades the rate rather than the run. It is deliberately not voice, not realtime, and runs no code execution by default, and the method tags every workflow step as AI (executable), DECISION (human gate), or TYPING. So the honest way to score these roles is: what percentage of each job's stated workload maps to AI-taggable steps a governed text agent can execute, with the human owner kept at the DECISION gates.

Link status. Three Wellfound postings (3829989, 4053494, 4004430) now return 410 errors — they've been taken down, so I can't score them. The last four links (workatastartup.com/jobs, the Wellfound role page, topstartups.io, startup.jobs) are job-board directory pages, not individual postings, so there's no single role to evaluate there.

Suitability table

The structural caveat that applies to every row

None of these scores can approach 100%, and that's by design rather than a capability gap. All eight are founding roles — the postings repeatedly ask for architecture ownership, product judgment, customer conversations, ambiguity tolerance, and company-shaping. Under the record's own counting rules, those are DECISION rows, and the method deliberately keeps a human at every one of them: spec approval, plan review, governance questions, and the first deploy all stop for the human's word. So the realistic deployment isn't "Raza-SOR takes the job"; it's "Raza-SOR executes 20–65% of the job's workflow (the build-agents, write-evals, wire-harness, draft-spec rows) while you or another human holds the founding-engineer judgment seat." The best-fit postings (Astoria, Allus, Vector Legal) are the ones whose day-to-day task lists are closest to the ten-step build loop the record already encodes.

If you'd like, I can take the top two or three and break each into a proper delegation map — step-by-step AI/DECISION/TYPING tags with the two counts — which would replace these estimated percentages with counted ones the way the method intends.

Raza-SOR is most suitable for supporting applied-agent engineering at Allus AI and Astoria AI. It is not currently a complete replacement for any of these founding-engineer roles.

I reviewed all 11 job descriptions, recovering indexed descriptions where direct Wellfound links failed. I also checked your Build Your Agent skill v2.0 and SOR generation 20. This assessment concerns an AI employee’s capabilities, not your suitability as a human applicant.

One important distinction: your skill guides a coding agent through specifications, implementation, tests and evaluations. The basic worker it creates is text-based and explicitly excludes code execution. It does not automatically inherit Codex’s ability to edit repositories, run commands or deploy software. Current SOR scope

For the comparison below, I assume the employee has a capable coding runtime, repository access, an isolated test environment and approved company integrations alongside SOR and the skill. These are required additions, not capabilities supplied by the SOR connection alone.

• Suitability: estimated fit across the role, including work requiring human review.
• Autonomous share: estimated responsibilities it could execute within an approved scope without continuous guidance. Final approvals still apply.
• These are judgment-based ranges, not measured results, time savings or employee-replacement percentages. The postings do not specify how much time each responsibility takes.

The three recovered Wellfound descriptions were PunttAI, Glacis and SureBright. Their descriptions could be assessed, but successful retrieval of indexed content does not confirm that applications remain open.

The remaining four links are directories, so they have no single responsibility set to score:

My recommendation: position the offering as an AI engineering worker that delivers a defined part of a team’s workload. Allus AI and Astoria AI provide the closest task matches. Vector Legal also offers useful scoped opportunities when domain experts supply and review the evaluation data.

For example, a credible first assignment would be:

“Given your approved specification and test data, implement one tool-using workflow, prepare regression tests, investigate failures and deliver a reviewable change.”

To establish an actual suitability percentage, test it on representative company tasks and record accepted outputs, human correction time, failures and unauthorized-action attempts. Completing a demo once would not establish that it can own the whole role.