Agent plugin operations

Agent plugins should make safe workflows easier to use, not hide unclear permissions or unsupported side effects.

When to use it#

Use this guidance when a developer tool, AI agent, or customer-facing assistant needs controlled access to systems, documentation, code, data, or workflow actions.

Operating boundary#

Boundary	Assistance can operate	Customer owns
Tools	MCP server runtime, tool schemas, connector deployment, health checks, audit logs.	Which systems may be exposed and which actions are allowed.
Access	Service accounts, secret rotation procedure, least-privilege recommendations.	User approvals, identity source of truth, and periodic access review.
Quality	Eval cases, regression checks, rollout gates, trace review.	Business acceptance and final approval for automated actions.
Support	Runbooks, dashboards, incident triage, change requests.	Business impact decisions and customer communications unless contracted.

Safe-use requirements#

Start with read-only tools unless the workflow explicitly needs mutation.
Scope tools by environment, tenant, user role, and business action.
Record tool calls, arguments, approval state, result, latency, and cost where relevant.
Keep prompts, skills, tool schemas, and server configuration versioned.
Define rollback and disable procedures before production rollout.

Onboarding inputs#

Bring the target workflow, systems to connect, data classification, allowed actions, identity model, environments, rate limits, expected users, and support contacts.

Handoff artifacts#

Assistance handoff normally includes the tool inventory, credential and rotation notes, server runbook, trace dashboard, evaluation evidence, known limitations, and change-request process.