Agentic AI explained: when an AI loops on its own decisions.
A normal LLM answers your question. An agent runs in a loop: pick a tool, run it, read the result, decide next step. Useful in narrow domains, dangerous when unsupervised.
9 min read
An agent is an LLM that decides what tools to call and in what order, runs them, reads results, and decides what to do next. The loop is what makes it an agent.
Why the loop matters
A single LLM call has fixed scope: you ask, it answers. An agent operates more like a junior employee: pick the right tool, observe the result, reason about whether the goal is met, choose the next step.
Example task: find the contracting officer email for a solicitation and draft a clarification question. A single LLM call cannot do this (no web access, no email tools). An agent can: search SAM.gov via a tool, parse the result, draft an email via another tool, return for human review. Five steps. Executed without human intervention.
Where agents work well
- Research workflows (searching, summarizing, comparing across sources) - Code refactoring at scope (read file, find callers, plan, execute, test, iterate) - Data wrangling (schema inspection, sample queries) - Document generation with structured intermediate steps
These share a property: the cost of a wrong intermediate step is low (re-runnable) and the final output is checkable by a human.
Where agents fail
- Anything with real-world side effects that cannot be undone. Sending money. Sending email to a customer. Deleting data. - Long horizons with many decisions. Each decision has ~5% error. Compound across 30 decisions and the agent is wrong 78% of the time. - Domains where success is hard to verify without expert judgment. Legal advice. Medical diagnosis. Compliance attestations.
How to deploy safely
1. Bounded scope. Single project, single dataset, single account. 2. Tool allowlist. Agent can call N specific tools, not "anything." 3. Human-in-the-loop on irreversible actions. Agent drafts; human approves and sends. 4. Observability. Log every tool call. Replay-able audit trail. 5. Budget. Token + time + tool-call caps. An agent not finishing in 30 calls is wrong.
The honest 2026 state
Agents work for research, refactoring, and structured intermediate-step generation. They do NOT work as fully autonomous employees. The buyer's job is to design the agent's scope so it cannot make the worst mistakes, even when it tries.
Apache-3 Inc.'s AI Operations Support helps agencies design these scope boundaries.
Related articles
Claude vs ChatGPT vs Copilot vs Gemini: when to use which.
All four are useful. They are not interchangeable. Picking the right one for the task saves time and money.
6 min read →
AI hallucinations: what they are, why they happen, how to spot one.
The word "hallucination" sounds dramatic. The reality is more mundane and more useful to understand.
6 min read →
AI vs ML vs LLM vs GenAI: the terms, in plain English.
Marketing departments use these words interchangeably. They are not interchangeable. Here is what each one actually means.
5 min read →