The architecture is the same everywhere. What changes is the tier, the tooling and who sits in the loop. Four patterns cover most of what enterprises are building right now.
What it does. Picks up a ticket, reads the codebase, proposes a change, writes tests, opens a pull request. The highest-value and fastest-moving pattern in most organizations right now.
Tier. R2 for pull requests into shared branches. R3 only for low-risk, well-tested repositories with mandatory review and automatic revert.
Tools. An agentic coding assistant with repository-scoped access, an MCP layer for the issue tracker, CI system and internal documentation, ephemeral build sandboxes with no production credentials, a secret scanner, an SBOM step for anything the agent introduces, and trace capture on every run.
Process. Scoped access per repository. Generated code flagged with provenance. Human review is mandatory and cannot be performed by another agent. Coverage and other quality gates apply exactly as they do to human-authored changes, with the same thresholds the repository already enforces. Cost per merged pull request is tracked as a first-class metric alongside cycle time.
People. Engineering lead owns the agent. Platform team owns the sandbox and the gateway. Security reviews the tool scopes quarterly. Reviewers rotate so no one person becomes the rubber stamp.
Watch for. Review fatigue. When pull request volume triples, review quality drops before anyone notices. Cap the number of open agent-authored pull requests per reviewer.
What it does. Handles a customer request end to end: understands the ask, retrieves the case, checks eligibility, takes the action or prepares it for a person.
Tier. R1 for information. R2 for anything that changes a customer record or commits money. R3 only inside a hard value ceiling with automatic reversal.
Tools. Orchestration on a durable workflow engine, MCP servers over the core systems exposing business operations rather than raw APIs, permission-filtered retrieval over policy and product documents, an ontology for product and entitlement terms, master data for customer identity, and a human takeover console.
Process. Value limits and tenant scope in the action envelope. Every customer-affecting decision logged with its evidence. Outcome monitoring across customer segments. A clear disclosure that an agent is involved, and a route to a person on request.
People. Operations lead owns the agent. Compliance signs the tier and the disclosure language. Frontline staff hold the takeover console and the escalation path. A quality team samples transcripts weekly.
Watch for. Silent scope creep. A tool added for one journey becomes reachable from every journey. Re-run the entitlement review whenever a tool is added.
What it does. Researches across internal and external sources, synthesizes an answer with citations, and drafts the document a specialist would otherwise write from scratch.
Tier. R1 almost always. The output informs a decision that a qualified person still makes and signs.
Tools. Hybrid retrieval with chunk-level permissions, a knowledge graph for entity resolution across sources, a citation enforcement step that rejects unsupported claims, subagents for parallel research with condensed returns, and versioned prompts per document type.
Process. Every claim traceable to a source or it does not ship. Retention aligned to the record-keeping rules of the domain. Reviewer attestation captured with the document. Regression set built from documents specialists have already approved.
People. A domain specialist owns quality and defines the golden set. A knowledge or data steward owns the sources and their classification. Legal defines retention and disclosure. Architecture owns the retrieval design.
Watch for. Confident synthesis across sources of different vintage. Timestamp every source in the context and require the agent to surface conflicts rather than resolve them quietly.
What it does. Triages alerts, diagnoses pipeline failures, proposes remediation, drafts the fix, and in mature setups applies a bounded class of routine changes.
Tier. R2 for anything touching production. R3 for a named list of routine, reversible operations inside a change window.
Tools. Read access to telemetry, logs and lineage. Write access only through the existing change pipeline, never directly to infrastructure. Runbooks exposed as tools. A guardian agent watching action sequences. Automatic rollback bound to every change class.
Process. The agent proposes into the same change process a person would use. Blast radius declared per action class. Change freezes apply to agents exactly as they apply to people. Every intervention is reviewed at the operations meeting.
People. The site reliability or platform lead owns the agent. The on-call engineer holds the kill switch. Data owners approve any change touching their pipelines.
Watch for. Cascading remediation. An agent fixing a symptom can create the condition for the next alert. Circuit-break on repeated action within a window.
Accountable for what the agent does. Approves the tier, holds the budget, answers for incidents. Named in the registry.
Owns the reference architecture, the standards register and the golden path. Decides where controls live and keeps patterns consistent across teams.
Builds the agent, its tools, its prompts and its evals. Works on the golden path by default.
Owns the golden sets, the thresholds and the judge calibration. Signs the release gate.
Owns the threat model, the tool scopes, the red-team cadence and the identity model for agents.
Owns the control mapping, the audit evidence and the regulatory interpretation. Engaged at design rather than at the gate.
Owns classification, retention and what an agent is permitted to know. Approves retrieval scopes.
Holds the kill switch, runs the incident drills, monitors cost and quality trends in production.
If you are standing up an agent platform, tightening the controls on one you already have, or preparing for an audit that now includes agents, I am happy to look at it with you.
Tell me where you are with agents and what you are trying to make safe. I reply to every enquiry within two business days.
