How Teams Actually Ship AI Agents
Methodology, harnesses, and governance for platform teams — written by practitioners, not vendors. Every recommendation names its failure mode; every comparison names the case where the losing option wins.
Guides
Prompt-injection defense in production: what actually held up in 2026
Six prompt-injection defenses tested against real 2026 attacks: what held, what failed, and the control your security team will demand before signing off on agentic deployments.
The Enterprise Agent Reference Architecture, Annotated
The annotated enterprise agent reference architecture: ingestion, orchestration, tool layer, eval harness, and governance plane - mapped to the decisions your platform team will actually face in production.
The Agent Platform RFP Scorecard
A scored RFP scorecard for evaluating AI agent platforms - covering harness architecture, eval maturity, governance controls, and vendor lock-in risk - so your security and finance teams can sign off.
What 2026 Enterprise Agent Deployments Have in Common
Six patterns that show up in every named public enterprise agent deployment in 2026 - extracted from the case studies, earnings calls, and vendor disclosures that are actually on the record.
Data Boundaries for Agents: RAG vs Fine-Tune vs Context Windows, a DLP View
RAG, fine-tuning, and long-context windows each create a distinct data-access path, and with it a distinct exfiltration surface. This maps all three through a DLP lens so your security team knows which controls to place where.
The EU AI Act GPAI Obligations, Mapped to Your Agent Stack
GPAI enforcement went live August 2, 2026. Here is what the EU AI Act actually requires of enterprise teams deploying agents: which obligations land on you vs. your model provider, what auditors will ask for, and the three controls that close most gaps.
Writing Agent PRDs: Acceptance Evals as the Contract
How to write an agent PRD so that every acceptance criterion has a direct eval counterpart - the document structure, the five fields each requirement needs, and what breaks first when the PRD and the eval suite drift.
Build vs Buy for Agent Orchestration: When Temporal-Style Beats Managed
A decision framework for platform teams choosing between managed agent orchestration (Bedrock, Vertex, Anthropic ODE) and self-hosted durable execution (Temporal, Inngest, Conductor). With the flip conditions, real cost math, and what breaks first.
AI Coding Assistants at 500 Seats: Rollout, Measurement, and the Renewal Conversation
A rollout framework for platform leads deploying AI coding assistants at 500 seats: which plan to pick, what to measure from day one, and how to own the renewal conversation with data.
Pilot, Platform Team, or CoE: the Three AI Agent Org Designs That Work
The three org designs for an AI agent function, keyed to headcount and governance: embedded champions under 15 engineers, a platform team to 150, a CoE above it, and the transition that breaks first.
CI for agents: gating merges on eval scores without blocking every PR
A tiered eval-gate design for agent pipelines: deterministic checks every commit, LLM-judge on merge, regression nightly. Worked GitHub Actions config, dated judge-cost math, and what breaks first.
Token-Budget Modeling for Platform Teams: a Spreadsheet Your CFO Can Audit
The cost model to build before an AI-tool rollout, not after the bill lands: four cost variables, a worked 50-to-150-seat token budget priced from live Anthropic API rates, and the five questions your CFO will ask.
Observability for Agents: OpenTelemetry GenAI Conventions in Practice
What the OpenTelemetry GenAI semantic conventions cover, what they still leave to you, and a two-hour setup that gets a platform team from a single duration metric to span-level agent traces.
Agent Sandboxing Patterns That Survive a Pentest
The isolation patterns for production AI agents that hold up when someone actually attacks them, and the specific failure each control does not stop. Threat model, isolation ladder, credential boundary, egress policy, and what your security team will ask.
The Agent-Skill Security Review: A Checklist Your Sec Team Will Sign
A 12-point security review checklist for AI agent skill files, mapped to the org-policy artifacts your security team needs to sign off. Builds on SkillSpector's 68 vulnerability patterns and the Hugging Face agent intrusion disclosure.
MCP in the Enterprise: Governing Tool Servers Before the Standard War Settles
Enterprise-Managed Authorization reached stable in June 2026, but MCP governance goes deeper than OAuth. Here is the vendor-vetting process, scope policy, monitoring checklist, and what your security team will ask before they sign off on any MCP tool server.
Eval Harnesses Compared: Braintrust vs LangSmith vs Promptfoo vs Arize Phoenix
Four LLM eval harnesses priced and compared on the axes an engineering lead actually decides on: trace capture, CI gating, self-host, and the cost to switch.
The Enterprise Agent Harness Decision: Claude Agent SDK vs OpenAI Agents SDK vs Build-Your-Own
The agent-runtime decision does not turn on token price or a feature checklist. It turns on who runs the durable-execution substrate and how much platform headcount you have. Priced three ways, by org profile, with the escape cost of each.
How Enterprises Actually Ship AI Agents in 2026
The path from shadow AI to a governed agent capability, staged. The maturity model, the reference architecture, who owns it, the real seat-vs-token economics, and the failure modes that break a rollout first.
Spec-Driven Agent Development: From PRD to Passing Evals
Write the agent spec first, derive acceptance evals directly from it, then gate merges on those evals. A worked PR-review example from spec to a green CI gate, plus what rots first.