Engineering questions on building AI agents

Anthropic draws the line cleanly. A workflow follows a predefined path; an agent decides the path at runtime. In production this matters because workflows fail predictably, agents fail unpredictably. Most production-grade systems are workflows wearing some agent affordances. We use the word 'agent' specifically when the system can choose its next tool, model, or sub-agent at runtime.

No, and Anthropic explicitly recommends starting with direct API calls. Frameworks add abstraction that helps once you have multi-pattern composition or persistent state, and adds overhead until you do. Our practice: hand-write the first prototype. Adopt a framework only after the orchestration shape is stable. We use LangGraph in production and tried CrewAI for two months.

There is no universal answer. We use Claude Code for terminal-based, multi-file work and Cursor for IDE-bound editing. Devin shines on isolated tasks where you want a managed sandbox, and is the most expensive when its planning prompt drifts. The honest comparison rule lives on our claude-code-vs-cursor and devin-vs-claude-code pages.

Yes for complex multi-file tasks, no for simple completion. The trick is the structure you put around it before you hit run, not the prompt at run-time. Used unsupervised it is the most expensive coding agent we have used; gated behind a planning step it is the cheapest per resolved task.

Prompt chaining, capped at three steps. Each additional step compounds drift super-linearly while compounding cost only linearly, so the cheap-window is short. Beyond three steps the routing pattern wins on cost because routing serialises only the model that needs the work.

Three reasons. The +9,999,900% YoY signal reflects a low base, not a saturated category. The rebrand history (clawdbot, moltbot) consolidated brand awareness onto a single name. The NVIDIA NemoClaw partnership validated the architecture for engineers reading hardware-vendor blogs. Whether it is safe to run in production is a different question, covered in the OpenClaw review.

Pick LangGraph if you expect to scale; pick CrewAI if you expect to ship fast. We use LangGraph in production. We tried CrewAI for two months and moved off because, around five concurrent agents, the role-based coordination overhead becomes the bottleneck. Your project may not hit that ceiling. If it will not, CrewAI is the faster path to a working prototype, and that is genuinely valuable.

Five categories, ranked by frequency in our pipeline. Silent drift (agent completes the task, the task is wrong, no exception). Cost cliff (orchestrator-worker spike). Confidence-gate breach (routing fails on a boundary case). Tool-call cascade (tool returns unexpected schema). Context-window blow-out (long-running agent exhausts context, hallucinates from the truncation point).

Past five, the orchestration overhead starts to dominate the work. Past eight, in our experience, the synchronisation cost exceeds the LLM token cost. This is a measured finding from running orchestrator-worker patterns across our pipeline. Most teams hit this ceiling sooner than they expect.

Depends on the isolation model. OpenClaw in a sandboxed configuration is operationally similar to running any third-party binary; in a non-sandboxed configuration, the security caveats are real and material. Self-hostable does not mean self-isolating. The OpenClaw review documents what we observed in both modes.

'Agentic' describes a property: the system can take actions toward a goal with some degree of decision-making latitude. 'Autonomous' describes a degree on that property: how much latitude. A workflow with one routing branch is mildly agentic and not autonomous. A loop that selects its own tools without human approval is highly agentic and approaching autonomous. The cornerstone definition lives on whatisanaiagent.com.

Skip the frameworks. Write a single Python script that calls a model API, parses the response, and calls one tool. Read Anthropic's 'Building Effective Agents' paper. Then read our prompt-chaining essay. Then add a second tool. Build up to a routing pattern. Adopt a framework only when you outgrow the script. The Maturity Curve names the stages.