Twinkle Savani
MenuClose
Twinkle Savani

ThreadsColophonContactLinkedInGitHub

Experimentexperimental

Orka

An autonomous SDLC orchestrator: take a GitHub issue, interrogate it, plan it, and open a tested pull request — with human approval kept explicit.

The code is private for now; the loop is exactly what it runs.

Orka is an experiment in taking the whole delivery loop, rather than the coding step, and handing it to agents. The first product loop is deliberately narrow:

  1. ingest a GitHub issue
  2. validate the requirements
  3. ask clarifying questions when they’re underspecified
  4. plan the implementation
  5. generate a tested pull request
  6. stop and ask for an explicit approval on anything risky

Rails 8, Sidekiq, Postgres, Hotwire, RSpec.

IssueValidateClarifyPlanTested PRApprovestops when underspecifiedhuman, explicit
The six steps, as built. Steps three and six are the ones that decide whether the loop is usable.
ai/agents/registry.ymlglobal rules, as written
  1. AI must never assume missing requirements.
  2. High-risk changes require founder approval.
  3. Every agent action should be observable, explainable, and reviewable.
  4. Prefer small pull requests and incremental commits.
  5. All agent handoffs go through one shared payload — no direct agent-to-agent calls.

Nine agents read these before anything else: product owner, delivery manager, designer, solution architect, software engineer, automation, security reviewer, code reviewer, founder.

ai/prompts/templates/readiness_report.mdwhat step two produces

Requirement readiness report

Summary · user story · acceptance criteria · clarification questions · readiness score 0–100 · risks · approval gates · recommended next agent

Blocker verdict: ready / needs clarification / blocked. Nothing is planned, let alone written, until this says ready.

The interesting part isn’t step 5. Generating code is the part that already works. The parts that decide whether the loop is usable are steps 2, 3 and 6 — whether the system can tell that a ticket is underspecified before it writes anything, and whether a human stays genuinely in the loop where it matters instead of nominally in the loop everywhere.

agentsexperimentsproductprocess