文章

共 30 篇。生产级 AI 智能体、工作流自动化与运营系统。

AI Agent Security: Threats That Matter in OpsPrompt injection is real. So is a service account that can wire money because someone gave the agent finance API scopes. Practical notes for teams shipping o…
MCP for Enterprise Agents: Inventory Before You ConnectMCP makes tools easy to attach. Easy attachment is how shadow IT becomes shadow toolchains. Practical notes for teams shipping operational AI agents.
Multi-Agent Systems in the Enterprise: When More Agents HelpsMore agents is not a strategy. Clear ownership of steps is. Practical notes for teams shipping operational AI agents.
Agentic Workflows vs RPA: What Actually ChangesRPA clicks UIs. Agents should prefer APIs and structured documents—and still fail loudly. Practical notes for teams shipping operational AI agents.
AI Agent Memory: What to Persist (and What Not To)Memory is state. If you would not store it in your CRM, do not hide it in an opaque vector bag. Practical notes for teams shipping operational AI agents.
AI Agent Permissions: Role Design That Matches HumansIf no human role has that power, your agent should not either. Practical notes for teams shipping operational AI agents.
Durable Execution for AI Agents: Why Crashes HappenAgents wait on humans and APIs. If your runtime dies mid-wait, the case should not die with it. Practical notes for teams shipping operational AI agents.
Grounding AI Agents in Real Systems (Not Vibes)Grounding means the agent’s inputs and outputs bind to systems of record—not that the prose sounds confident. Practical notes for teams shipping operational …
Human-in-the-Loop AI Agents: Design That Survives ProductionA human in the loop is not a disclaimer. It is a product surface: queue, context, SLA, and override reason. Practical notes for teams shipping operational AI…
AI Agent Observability: What to Log in ProductionIf you cannot reconstruct why a case closed, you do not have a production agent—you have a demo with uptime. Practical notes for teams shipping operational A…
AI Agent Rollback: When the Model Was WrongIf you cannot undo or compensate a bad write, do not let the agent write unsupervised. Practical notes for teams shipping operational AI agents.
AI Workflow Automation: Start With One PathAutomation programs that start with "transform all ops" finish with a pilot graveyard. Start with one path. Practical notes for teams shipping operational AI…
Exception Queues for AI WorkflowsAutomation without an exception queue is a firehose aimed at whoever yells last. Practical notes for teams shipping operational AI agents.
Document AI Agents: Extraction Is Not the ProductExtracting text is table stakes. Completeness rules and system write-back are the product. Practical notes for teams shipping operational AI agents.
Tool-Calling Agents in the Enterprise: Permissions FirstTool calling without least privilege is remote code execution with a friendly UI. Practical notes for teams shipping operational AI agents.
AI for Ops Automation: Queues, Not Dashboards AloneOps leaders need fewer manual touches per case—not another dashboard that narrates the backlog. Practical notes for teams shipping operational AI agents.
Compliance AI Agents: Controls Over ChatCompliance agents should open findings with evidence—not declare you "compliant" in a paragraph. Practical notes for teams shipping operational AI agents.
RAG vs AI Agents: Different JobsRAG answers from your docs. Agents change state in your systems. Confusing them burns a quarter. Practical notes for teams shipping operational AI agents.
AI Agent Cost Control: Tokens Are Not the Whole BillToken spend is visible. Human review time and failed write-backs are the quiet budget killers. Practical notes for teams shipping operational AI agents.
AI Agent Evaluation: Beyond Demo Happy PathsIf your eval set is five clean PDFs, production will surprise you. That is not unlucky—it is under-tested. Practical notes for teams shipping operational AI …
Enterprise AI Agents in 2026: What Actually MaturedThe hype cycle moved. The production constraints did not: permissions, evals, and owners. Practical notes for teams shipping operational AI agents.
AI Agent Guardrails That Are More Than a System PromptA system prompt is not a control. Controls are code paths that cannot be talked out of. Practical notes for teams shipping operational AI agents.
Custom AI Agents vs Agent PlatformsPlatforms accelerate demos. Custom systems fit when policy and systems are the product. Practical notes for teams shipping operational AI agents.
Production AI Agents: A Practical ChecklistProduction-ready means someone on-call can explain the last failure. Everything else is marketing. Practical notes for teams shipping operational AI agents.
AI Agent Frameworks: Pick for Operations, Not HypeFramework choice is reversible if you keep schemas and tools behind your interfaces. Workflow design is not. Practical notes for teams shipping operational A…
AI Agent Orchestration: Control Plane, Not ChaosOrchestration is scheduling, state, and policy—not letting five models freestyle a plan. Practical notes for teams shipping operational AI agents.
Computer-Use Agents at Work: Caution FirstGUI-driving agents look like sci-fi RPA. They also inherit every UI fragility—and add model error. Practical notes for teams shipping operational AI agents.
Browser Agents for Enterprise Ops: Scope TightlyA browser agent is a tool. Without case state and permissions, it is a script with a probability distribution. Practical notes for teams shipping operational…
KYC AI Agents: Automation Without Auto-Clearing RiskKYC agents should prepare and check—not auto-clear high-risk customers because the model felt confident. Practical notes for teams shipping operational AI ag…
LLM Agents for the Enterprise: Buy, Build, or WaitEnterprise agent projects fail more from unclear workflows than from model choice. Practical notes for teams shipping operational AI agents.