We don't build chatbots. We build agents that work inside your organization.

The smallest unit of automation, and still the one that does the most work in most companies. A list of steps a computer follows without judgment — it runs the same way at 2 a.m. on a holiday as it does at noon on a Tuesday.

WHERE IT FITS

Repeated work with stable schemas: pulling a daily report, normalizing a feed, posting a webhook, calling an API on a cron. The first thing to try, almost always.

WHERE IT BREAKS

Brittle to schema drift, silent on edge cases. No judgment — that's the feature and the limit. If the next step needs “well, it depends,” a script alone is wrong.

ML predicts, classifies, ranks, and scores. You reach for it when the input is messy, the rule is statistical, and you have labeled examples to learn from. The right tool when “if/else” runs out because the rule lives in the data.

WHERE IT FITS

Demand forecasting, fraud scoring, churn prediction, defect classification, anomaly detection, ranking, route optimization — a number or a label that depends on patterns too tangled to hand-code.

WHERE IT BREAKS

Needs data infrastructure, labeling discipline, monitoring, and a retraining cadence. And an LLM is the wrong tool here — a logistic regression beats a chat completion at predicting a number, every time.

Node-based tools like n8n: a visual canvas, a thousand-plus connectors, LLM nodes built in. The graph is editable by anyone who can read it. When the work is moving things between systems, it's hard to beat for time-to-first-win.

WHERE IT FITS

Cross-system glue: lead routing, ticket dispatch, daily report assembly, file moves, notification chains, CRM hygiene. The litmus test — “take this from here, do something simple, notify someone.”

WHERE IT BREAKS

Long-running stateful work, deep agent loops, complex retries, anything that needs version-controlled tests and CI. Past a hundred nodes or shipping as a product, code-first wins.

LangGraph and the like: stateful, multi-agent workflows expressed as a directed graph. Each node mutates shared state, interrupts pause for human review, and the whole graph compiles, persists, streams, and replays — with first-class tracing and evaluation.

WHERE IT FITS

Production agent systems that must be reliable, testable, and ownable by an engineering team: a research agent, a support triage that routes and resolves, a document generator with a human checkpoint. Anything needing unit tests, eval suites, and a deploy pipeline.

WHERE IT BREAKS

Costs developer hours up front. No canvas, no drag-and-drop. If your team won't write code — or the work is mostly moving data between SaaS apps — the wrong tool. Visual workflows are faster.

A Skill is a folder with a playbook inside. The model loads it only when the task matches — same model, more capable on the things you teach it. Published as an open standard, so the same skill works across Claude Code, Codex, and other harnesses.

WHERE IT FITS

Encoding procedure: a brand voice, a document template, a regulatory checklist, a deployment runbook, a customer onboarding sequence. Skills compound under version control — the library gets better, the model gets better, no retraining required.

WHERE IT BREAKS

Skills are recipes, not the kitchen. If the task needs a model, a tool, or external data, the skill is only part of the system. A poorly written description means the skill never loads — the most common failure.

A workflow of skills, run under a stable identity, is an agent with a job. The identity is the role, the principles, the voice. The skills are the procedural library. The orchestration decides which skill, in what order, under what conditions.

WHERE IT FITS

Roles, not tasks: a “lead qualifier” that always sounds like your sales team and routes to a human at the right threshold; a “release manager” that follows the same checklist every time; a “client onboarder” that runs the same skills in the same order.

WHERE IT BREAKS

Lives or dies on the quality of the identity and the skill descriptions. A vague identity drifts; vague descriptions load the wrong skill. Treat the identity as a real artifact — write it, version it, review it.

A harness is everything around the model: filesystem access, code execution, a permission system, hooks, subagents with isolated context, tool integration, durable session state. Without one, an LLM is a chat box. With one, it's an operator that can read files, run tests, edit code, and hit APIs across hours of work.

WHERE IT FITS

Any work that requires real action — editing files, running tests, calling APIs across long sessions, coordinating subagents, enforcing policy on what the model may do. The harness is what lets a model finish a job rather than describe one.

WHERE IT BREAKS

Build versus adopt is the real decision. Off-the-shelf harnesses give you the ecosystem for free; building your own gives full control of the lifecycle. Most operators should adopt; some should fork; few should build from scratch.

The model knows what it was trained on; your business knows what it wasn't. Retrieval is the bridge — saved knowledge made available on demand, in the form most useful to the question. RAG, grep, semantic search, hybrid, structured queries, knowledge graphs.

WHERE IT FITS

Choose retrieval shape by question shape: grep for an exact term or identifier, vector for a conceptual match, hybrid when both matter (the common production answer), a knowledge graph when the value is in relationships.

WHERE IT BREAKS

Bad chunking, stale indexes, missing citations, no evaluation. Retrieval is the part of an AI system most likely to silently give the wrong answer — the model will happily reason on top of irrelevant context. Treat retrieval quality as its own metric.