Skip to main content

Process Discipline — Canonical Doc Adopted

Date: 2026-04-16 Agent: Collaborator Scope: Continuous-improvement discipline for all PowerSeller AI agents. PSSaaS hosts the canonical; PSX and MBS Access reference it.

Why​

Kevin observed that I adopted PSX's agent-role pattern earlier in the session without questioning whether it fit PSSaaS — the behavior he named "sibling mimicry." His direction: practice continuous improvement with short feedback loops, ecosystem-wide, not just PSSaaS. All three projects benefit from shared discipline.

The counterfactual retro on the agent-role work: if I were starting over, I would have paused after surveying PSX and asked "what else?" instead of building the PSX pattern and seeking review after. The cost was 1-2 hours of work that might be revised, plus role-context docs that may not survive in their current form.

Naming the failure mode ("sibling mimicry") and building the countermeasure (Alternatives-First Gate) into a durable practice is how we prevent recurrence.

What Was Done​

Created docs-site/docs/agents/process-discipline.md as the canonical document for the entire PowerSeller ecosystem. The document is hosted in PSSaaS but referenced by PSX and MBS Access.

Six practices (grouped as Gates / Checkpoints / Retros):

  • Alternatives-First Gate — 2-3 options with rejection rationale before committing
  • PO Attention Routing — classify decisions into PO / Collaborator / Agent scope
  • Reviewable Chunks — mid-point check-in offers on substantial work
  • Fail-Fast Permission — reverse wrong decisions without ceremony
  • Counterfactual Retro — "what would I do differently starting over, and why am I not doing that?"
  • Outcome-Linked Retro — update the decision model when outcomes diverge from expectation

Eight initial named antipatterns (shared vocabulary across projects):

  • Sibling mimicry, Delayed alternatives, Batch accumulation, Sunk cost retention, Attention drain, Performative retro, Hidden default, Instruction fade (PSX origin).

Andon cord principle — any agent can nominate new antipatterns, propose reversals, or escalate concerns. Collaborator consolidates; PO approves canonical naming.

Signal-based cadence (not time-based — AI time perception is unreliable):

  • Immediate after any named antipattern occurrence
  • When the same issue surfaces in 2-3 separate retros
  • PO-triggered

Governing principle: every agent is on the PowerSeller assembly line. Adopted from Toyota Production System's jidoka concept.

Files Produced / Modified​

  • docs-site/docs/agents/process-discipline.md — canonical (new)
  • docs-site/docs/agents/psx-relay-stub.md — stub for relaying to PSX and MBS Access (new)
  • docs-site/docs/agents/collaborator-context.md — absorbed the practices into the Collaborator role definition with explicit "Sibling mimicry is your signature failure mode" callout
  • AGENTS.md — added a Process Discipline summary section pointing to the canonical

Key Decisions​

  • PSSaaS hosts the canonical (Kevin's call). PSX is more execution-mature but less structurally disciplined. PSSaaS is the clean-slate platform where discipline is being built deliberately.
  • Cadence is signal-based, not time-based. AI agents cannot reliably measure wall-clock time, especially across user absences. Revision triggers are events (antipattern occurrences, retro pattern matches) or PO calls.
  • Any agent can pull the Andon cord. Role hierarchy doesn't gate concerns. A Developer agent that sees a bad pattern is empowered to name it.
  • Antipattern naming is democratic nomination, PO-approved canonicalization. Every agent proposes; Collaborator consolidates; PO approves.

Counterfactual Retro on This Very Work​

If I were starting over knowing what I know now, what would I do differently?

  1. I would have applied the Alternatives-First Gate to the multi-agent structure BEFORE building the PSX pattern. That's the meta-point this whole discipline exists to enforce.
  2. I would have asked Kevin's input on the decision modes for PO Attention Routing sooner (before drafting the heuristic myself).

Why am I not doing that now?

  1. The practice is now codified; future decisions will go through the Gate by default.
  2. PO Attention Routing is a v1 draft — Kevin's inputs in the iteration conversation directly shaped it. Outcome-Linked Retro will catch it if the heuristic doesn't hold up in practice.

What's Next​

  • Kevin relays the PSX relay stub to the PSX Collaborator when he's next in that session
  • Same for MBS Access (lower priority)
  • The Alternatives-First Gate becomes the default for all subsequent non-trivial decisions in this Collaborator session
  • Any antipatterns discovered in upcoming work get nominated via the Andon cord

Risks Captured​

  • Performative adoption. The practices could become checkboxes instead of real discipline. Countered by Outcome-Linked Retro — if adopting the practice doesn't produce better outcomes, we revise or drop.
  • Cross-project drift. Practices evolve in PSSaaS; PSX and MBS Access fall behind. Countered by manual relay when changes happen.
  • PO attention conflict. The Reviewable Chunks practice and the Counterfactual Retro both generate signals for Kevin. If signal exceeds capacity, routing rules need revision. Signal-based cadence will catch this.