guardrails

DE//AGENT currently lists 7 resources under guardrails: 4 tutorials, 2 patterns, 1 plugin. Everything below is derived from the live registry — counts, capabilities, verification state and security posture come from the curated records, never from marketing copy.

Documented learning objectives across these resources:

• Define a unit metric per pipeline — via Install cost guardrails before the invoice
• Baseline the last N runs and fail CI past a threshold — via Install cost guardrails before the invoice
• Add hard caps on top of CI guards — via Install cost guardrails before the invoice
• Separate agent read/recommend from write/execute — via Keep an agent read-only until a human approves
• Encode the safety policy as hooks — via Keep an agent read-only until a human approves
• Keep a human gate on anything irreversible — via Keep an agent read-only until a human approves
• Explain where hooks run — on tool-call and pre-run events, outside the model — via Hooks — deterministic gates around an agent
• Match each safety hook to the failure it prevents — via Hooks — deterministic gates around an agent

4 more objectives live on the individual resource pages.

These resources map to stages: Reliability (2), Agentic Data Engineering (5). Follow the four-phase roadmap to place this topic inside a progression instead of learning it in isolation.

Every entry, its shape and its trust state — sorted by the registry's composite score. This table is the honest version of a marketing comparison: no sponsored placements, no unverifiable claims.

How scores are computed: source quality score.

Highest composite score first:

1. Agent-in-the-Loop Review score 35.75 — intermediate
2. Install cost guardrails before the invoice score 34.15 — beginner
3. Keep an agent read-only until a human approves score 34.15 — beginner

Scores rank provenance, completeness and trust — read how verification works before adopting anything.

Start here if you're new: Install cost guardrails before the invoice, Keep an agent read-only until a human approves, DE Data Toolkit.

Security-flagged entries: DE Data Toolkit (2 flags), Cost Guardrails (3 flags). Flags mean the record declares capabilities like credential use or write access — read the security posture on each page before connecting.

Known limits: A purely local sandbox with no real data (Agent-in-the-Loop Review)

• Agent-in-the-Loop Review — Coding agents can do real data-engineering work, but can also ship a schema change or drop a table. details
• Install cost guardrails before the invoice — Instrument, attribute, then guard: the Cost Guardrails pattern in four moves — a unit metric per pipeline, a baseline of the last N runs, CI failing on regression, and hard caps. I… details
• Keep an agent read-only until a human approves — Coding agents can do real data-engineering work, but can also ship a schema change or drop a table. Walk the Agent-in-the-Loop Review pattern: separate read/recommend from write/ex… details
• Hooks — deterministic gates around an agent — The model can be talked into anything; a hook is code on the tool path. Walk the four safety hooks every data repo should know — destructive-SQL guard, prod write confirmation, sec… details

Real install commands from the registry — read the source before running any of them:

amp plugin add de-agent/de-data-toolkit

DE Data Toolkit — review first

Adjacent topics: tutorial (4 shared), review (4 shared), agentic (4 shared), cost (2 shared), architecture (2 shared), plugin (2 shared).

Study path: the roadmap places this topic in a progression · the scoring model explains the numbers · the trust rule catalog shows every security check these entries passed through.