AI Agent Development

What does an AI agent actually do in production?

A production agent receives a goal, picks from a fixed set of tools, calls them with validated arguments, inspects the result, then continues or escalates to a human. Every step is logged. The hard engineering is not the prompt. It is the tool contracts, the retry policy, and the stop conditions.

The prompt is maybe five percent of the work. The rest is the boring layer nobody demos: a schema for every tool argument so the model cannot pass a malformed record into your CRM, idempotency keys so a retry does not create the same deal twice, timeouts so one slow API does not stall a run, and a budget ceiling per task so a loop costs you dollars instead of a bill you find out about on Friday.

On Chasr, the useful part was never the model. It was fitting GPT-4 follow-up assistance inside an existing HubSpot workflow so reps did not have to change tools. Reps abandon anything that adds friction. The result was measured in recovered hours, not in a new dashboard: 10 hours saved per rep per week across 50+ reps, follow-ups up 60%, and $100K+ of capacity recovered per quarter.

When is an agent the wrong answer?

When the task has a known sequence. If a human would follow the same seven steps every time, write those seven steps in code and use the model only where judgment is required. Agents earn their cost on branching, ambiguous work. On fixed work they add latency, spend, and a new class of silent failure.

ScamMinder is the clean example. Deciding whether a domain is a scam sounds like agent work. It is not. It is 14 deterministic verification phases feeding signals into a single GPT-4o classification step. That design is why it holds sub-6-second analysis at 20,000+ domains and 99.9% uptime. An agent free to choose its own path would have been slower, more expensive, and impossible to explain to a user disputing a verdict.

My first deliverable on most engagements is a written recommendation on how much autonomy the system should actually have. Sometimes that recommendation is that you do not need an agent at all. That conversation costs you an hour and saves a quarter.

How do you stop an agent from looping or burning budget?

Hard limits, not good intentions. A maximum step count per run, a token and dollar ceiling per task, timeouts on every tool call, and a cheaper model for the steps that do not need a frontier one. Anything that exceeds a limit stops and escalates to a human with its full trace attached.

Cost control is an architecture decision, not a setting. Route each step to the cheapest model that clears its quality bar, cache anything deterministic, push long work into queued background jobs instead of holding an HTTP request open, and measure spend per completed task rather than per token. The AI SEO platform runs this shape at scale: 13 Inngest background pipelines, 8 CMS adapters behind one publishing interface, 100 API routes, and 74 test files.

Provider routing matters as much as model choice. Running LLM calls through OpenRouter means a provider outage or a price change is a config edit rather than an incident. Every run writes a trace, so when an output is wrong you can see which tool returned what, and fix the actual cause instead of rewriting the prompt and hoping.

What proof do you have that this works outside a demo?

Five systems using LLMs under real load: Chasr, ScamMinder, an AI SEO content platform, Continuity — a Belgian M&A marketplace with a Claude-powered matching engine, an OCR dataroom, and 4 registry API integrations — and an ongoing GTM signal-intelligence build for a B2B SaaS team. None of them is an autonomous multi-agent swarm, and I will not describe them as one.

The GTM build is the closest of those to agent work, so it is worth being exact about. It monitors 20-30 target organizations, interprets what it finds in business context rather than pattern-matching keywords, recommends who to contact and why, and improves from the reviewer's corrections in a review UI. Monitor, interpret, recommend, learn — that is agent-shaped. What it is not is autonomous: a person approves before anything leaves the building, and the engagement is ongoing rather than a finished deployment. The distinction is not pedantry. A human-gated recommender and a fleet of self-directed agents have different failure modes, different spend profiles, and completely different review burdens.

That distinction is the whole point of this page. Deep production experience with LLM orchestration, tool calling, retrieval, background pipelines, and cost control transfers directly to agent work. It is not the same claim as having run a fleet of self-directed agents in production, and a developer who blurs those two things will blur other things later.

If you are interviewing agencies for agent work, ask each one which specific system they shipped, how many steps it took autonomously, and what happened the first time it failed. The answers separate engineering from marketing quickly.