skills.rest
Zhachory1
Zhachory1COMMUNITY
Explore all skills by this author →
Website𝕏@zhachory1GitHub

backprop

Optimizes AI agent workflows by analyzing run telemetry, failure clusters, and A/B rollout results.

Updated Jul 8, 2026
Search Tags:#ai agents#retrospective#failure analysis#telemetry#workflow optimization#a/b testing#rollout
One-click install
npx skills add https://github.com/Zhachory1/autopraxis --skill backprop-zhachory1
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: backprop
Source: https://github.com/Zhachory1/autopraxis/tree/main/skills/backprop
Command: npx skills add https://github.com/Zhachory1/autopraxis --skill backprop-zhachory1

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? AI agent workflows degrade over time through recurring failures, rising cost and latency, and human rework, but teams lack a systematic way to measure and fix these issues. This Skill turns workflow improvement into a measured loop: ingest run telemetry, cluster failure modes, attribute root causes, and promote or roll back changes based on evidence. ## Core Features & Use Cases - Telemetry-Driven Analysis: Ingests run telemetry JSONL, agent-fleet journals, long-term memory MCP notes, code RAG, git history, and PR/CI outcomes to compute per-step metrics like latency, cost, failure rate, and human-edit rate. - Hypothesis-Based Improvement: Generates testable change hypotheses with expected metric movement and regression risk, then validates them through council review and shadow or A/B rollout before promotion. - Safe Rollout and Prune Governance: Supports shadow mode, A/B testing, promote-or-rollback decisions with changelogs, and governed skill retirement that protects rare-but-critical skills. - Use Case: After noticing a planning workflow has a rising human-edit rate, run backprop to cluster the failures across recent runs, identify a weak rubric as root cause, shadow-test a revised prompt, and promote it only if guardrail metrics hold. ## Quick Start Ask the agent to run backprop on the recent workflow runs to find recurring failure modes and propose a measured improvement with a shadow rollout plan.

Frequently Asked Questions about backprop

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I optimize an AI agent workflow based on past runs?▼

Run backprop against your workflow's telemetry JSONL files and run artifacts. It computes per-step metrics like latency, cost, failure rate, and human-edit rate, clusters recurring failure modes, and proposes measured improvements validated through shadow or A/B rollout.

What data sources does workflow telemetry analysis need?▼

Backprop uses run telemetry JSONL, agent-fleet journals and transcripts, long-term memory MCP notes, code RAG repo indexes, git history, and GitHub PR reviews and CI results. Sources are used in priority order, storing only summaries and pointers unless durable capture is approved.

When should I use shadow rollout versus A/B testing for workflow changes?▼

Prefer shadow mode first for high-risk changes, since it runs the candidate against baseline without affecting routing. Use A/B testing only when routing and metrics are fair, with comparable task assignment and guardrail metrics to avoid unfair comparisons.

Can backprop retire or remove unused workflow skills?▼

Retirement is possible only in deep mode with full council and human approval. It requires per-skill invocation counts from telemetry lifecycle, and safety, incident, security, or release-critical skills are never retired on absence alone since they are legitimately infrequent.

Why does workflow optimization fail with insufficient data?▼

Proposals based on one or two runs overfit to noise. Backprop enforces a minimum data threshold, stops on insufficient data with instrumentation recommendations, and requires at least five runs of lifecycle signal before raising skill-inventory add hypotheses.

Related Skills

Explore More Software Engineering Skills→
Cogni-AI-OUCogni-AI-OU
3

brainstorm-agent-runs

Diagnose GitHub Actions workflow runs for a Pull Request and generate an Ishikawa diagram.

Official
Intermediate
RocapineRocapine
2

bump-version

Coordinate multi-package version bumps and changelog updates for related SDKs.

Official
Advanced
zephyrflow97zephyrflow97

test-driven-development

Enforce a failing test before production code through the RED-GREEN-REFACTOR loop.

Community
Intermediate
wildbitcawildbitca
1

flutter-logging

Enforce a unified Flutter/Dart Logger for all application logs.

Official
Intermediate