Process evolution: observe-propose-shadow-promote lifecycle #418

New issue

Open

opened 2026-04-08 10:21:14 +00:00 by dev-bot · 0 comments

dev-bot commented

2026-04-08 10:21:14 +00:00

Collaborator

Problem

The agent system cannot adapt its own processes. Modifying live process is too risky — a wrong step means the factory cannot recover. But never changing process means rigidity accumulates over time.

Proposal

Introduce a safe process mutation path with four stages: observe → propose → shadow → promote.

1. Observe (autonomous)

Agents track process metrics: review wait times, failure rates, escalation frequency, empty review cycles. Data accumulates in the knowledge graph.

2. Propose (autonomous)

When an agent detects a pattern (e.g. "last 15 docs-only PRs had zero review comments, median wait time 4 hours"), it writes a structured RFC to process-rfcs/ with evidence and proposed change.

3. Shadow (human-gated)

Human reviews the RFC. If approved, the new process runs alongside the existing one for N cycles. Both produce outputs; results are compared.

4. Promote (human-gated)

After successful shadow period, the old process is retired and the new one becomes canonical.

Key constraint

Agents can do steps 1-2 autonomously. Steps 3-4 require human sign-off. This gives agents voice in process improvement without giving them the ability to break the factory.

Rationale

Modeled on SRE canary deployment practices (Google) and continuous improvement loops (Toyota kata). The system learns and adapts, but mutations are controlled and reversible.

## Problem The agent system cannot adapt its own processes. Modifying live process is too risky — a wrong step means the factory cannot recover. But never changing process means rigidity accumulates over time. ## Proposal Introduce a safe process mutation path with four stages: **observe → propose → shadow → promote**. ### 1. Observe (autonomous) Agents track process metrics: review wait times, failure rates, escalation frequency, empty review cycles. Data accumulates in the knowledge graph. ### 2. Propose (autonomous) When an agent detects a pattern (e.g. "last 15 docs-only PRs had zero review comments, median wait time 4 hours"), it writes a structured RFC to `process-rfcs/` with evidence and proposed change. ### 3. Shadow (human-gated) Human reviews the RFC. If approved, the new process runs *alongside* the existing one for N cycles. Both produce outputs; results are compared. ### 4. Promote (human-gated) After successful shadow period, the old process is retired and the new one becomes canonical. ## Key constraint Agents can do steps 1-2 autonomously. Steps 3-4 require human sign-off. This gives agents voice in process improvement without giving them the ability to break the factory. ## Rationale Modeled on SRE canary deployment practices (Google) and continuous improvement loops (Toyota kata). The system learns and adapts, but mutations are controlled and reversible.