The strongest part is treating the main Claude session as an orchestrator rather than a single super-tool. Assigning PM, SWE, QA, and on-call roles makes the workflow easier to reason about because each agent has a responsibility and a verification boundary. I also appreciated that you tested it across real projects instead of presenting an agent team as a diagram only.
The part about agents declaring things done too early - hit this constantly. Turns out the failure mode isn't the model, it's that the agent had no external accountability loop. Once I added a separate QA role that explicitly re-ran acceptance criteria before close, the premature completions mostly stopped.
The explicit specification step before any coding is the thing I'd underline twice. Wrote a spec as if explaining to a new hire, not as a prompt. Night and day difference in output quality. How are you handling context bleed when the PM agent's spec feeds directly into the SWE agent?
the role assignment framing is the right one - PM, SWE, QA, on-call maps cleanly to what I have been running.
the interesting failure mode I keep hitting: agents in the QA and on-call roles surface process gaps that the PM agent then has to account for in the next cycle. the feedback loop between roles is where the complexity lives, not in the individual agent tasks. curious how you handled inter-agent escalation across the five projects.
The strongest part is treating the main Claude session as an orchestrator rather than a single super-tool. Assigning PM, SWE, QA, and on-call roles makes the workflow easier to reason about because each agent has a responsibility and a verification boundary. I also appreciated that you tested it across real projects instead of presenting an agent team as a diagram only.
What’s most interesting here is not that AI agents can now write code.
It’s that they can replicate organizational workflow.
PM agents, SWE agents, QA agents, on-call agents —
this is no longer “AI as a tool.”
It’s AI simulating operational structure.
The real implication is that AI may replace organizational friction before it replaces human intelligence.
But that also creates a new problem:
execution can now scale faster than governance, review, and long-term judgment.
Did you set up a sandbox environment for your agents to test their code before commiting it? Is that what you used CI/CD for?
The part about agents declaring things done too early - hit this constantly. Turns out the failure mode isn't the model, it's that the agent had no external accountability loop. Once I added a separate QA role that explicitly re-ran acceptance criteria before close, the premature completions mostly stopped.
The explicit specification step before any coding is the thing I'd underline twice. Wrote a spec as if explaining to a new hire, not as a prompt. Night and day difference in output quality. How are you handling context bleed when the PM agent's spec feeds directly into the SWE agent?
Perhaps adding a token budget for each Agent
the role assignment framing is the right one - PM, SWE, QA, on-call maps cleanly to what I have been running.
the interesting failure mode I keep hitting: agents in the QA and on-call roles surface process gaps that the PM agent then has to account for in the next cycle. the feedback loop between roles is where the complexity lives, not in the individual agent tasks. curious how you handled inter-agent escalation across the five projects.