CCDV-P · Agent Fleet Operations for Software Delivery
Configuring coding agents across the whole organization
A company of 400 developers adopts a coding agent. Within three months, each team configured it its own way: some with full permissions for convenience, others connecting MCP servers pulled from the internet, some routing traffic through a personal account. Security doesn’t know what’s in use, finance doesn’t know what’s being spent, and when one team writes a great Skill nobody else can benefit. Organization-level configuration fixes this.
Managed settings and policies
Managed settings are configuration the organization sets and people can’t bypass. In Claude Code there’s a configuration hierarchy (from the managed policy, the strongest, to the user’s) and a managed-policy file deployed with device-management tools. What gets defined there:
- Permissions: which tools and commands are allowed, ask or are denied; a denial in the policy can’t be overridden from below.
- Model and account: which models are used and under which account or contract, so usage goes through the organization and not personal accounts.
- Organization hooks: deterministic controls (block destructive commands, secret scanning before writes) that always run.
- What local configuration each team may add and what it may not.
MCP allowlists
An MCP server is code and access. The MCP allowlist defines which servers may be used, or which are blocked, and is enforced by managed policy so individuals can’t add arbitrary servers (see the tool-poisoning risk in the security lesson). Approved servers are reviewed (who maintains them, what permissions they request, how they authenticate), pinned by version and published in an internal catalog.
Telemetry
Telemetry answers what’s used, what it costs and whether it works: usage and cost per team and project, change acceptance rate, denied tools, errors, sessions. Claude Code can export metrics and events via OpenTelemetry to the company’s observability platform. It carries the same privacy care as any log: you measure metadata and decide carefully what content to capture.
Gateways and routing
A routing gateway concentrates agent traffic toward model providers: central authentication (no keys handed out to teams), per-team limits and budgets, logging and cost attribution, routing to different models or to a cloud under the company’s agreements (Bedrock, Vertex AI) and a single egress point to control. Confirm in the documentation how to point the agent at a gateway or a cloud; configuration per platform isn’t identical.
Governing what gets shared
Sharing prompts, Skills and MCP servers across teams multiplies both value and risk. Governance sets a quality and safety bar:
- an internal repository or marketplace with approved assets, and which marketplaces may be installed from;
- review before publishing: security (what it executes, what permissions it asks for, whether it reads sensitive data), quality (does it have evaluation and examples?), documentation and an owner;
- versioning and an update and retirement process, and a record of who uses what;
- the rule that a shared asset doesn’t widen permissions by itself: what it requests is compared with the organization’s policy.
Without this, one team’s Skill or server becomes a door nobody reviewed for the other 399.
The usual traps
- Per-person configuration, with no baseline policy.
- Full permissions “to be agile” with no organizational controls.
- Each team with its own key or personal account.
- Installing any MCP server or Skill from the internet without review.
- Measuring usage without measuring cost or outcomes.
On the exam
If a requirement must always hold for everyone, it goes in managed settings, not in a team guideline. If untrusted servers are the worry: allowlist and a reviewed catalog. If cost and attribution are the worry: gateway and telemetry.
Quick question
A practice exam question on this domain. It counts towards your progress.
A company wants a Claude agent to respond automatically when a production deployment pipeline fails. It is triggered by the pipeline's failure event, reads logs and recent commits, and investigates the cause. The first proposal lets the agent apply a fix and redeploy by itself whenever it is confident, to reduce the time to recovery.
The on-call engineers are concerned about two things: duplicate events during one outage that could launch several agents at once, and actions on production that nobody has reviewed.
Which design is the best fit for this event-driven responder?