|
| 1 | +--- |
| 2 | +title: "AI basics – LLM agents" |
| 3 | +description: "LLM-based AI agents: beyond chat, planning, memory, tools, ReAct, multi-agent patterns" |
| 4 | +date: "2026-03-25" |
| 5 | +slug: "ai-basics-agents" |
| 6 | +tags: |
| 7 | + - Artificial Intelligence |
| 8 | + - Machine Learning |
| 9 | + - База |
| 10 | +image: cover.jpg |
| 11 | +--- |
| 12 | + |
| 13 | +A «Bare minimum» article on **AI agents** built around language models: systems that **plan**, **retain context**, **call tools**, and iterate toward a goal—not just emit a single reply. |
| 14 | + |
| 15 | +### Chat vs agent {.toc-heading-only} |
| 16 | + |
| 17 | +<details class="post-accordion"> |
| 18 | +<summary style="cursor: pointer; font-weight: 600;">Chat vs agent</summary> |
| 19 | +<div style="margin-top: 0.75em;"> |
| 20 | + |
| 21 | +<p><strong>A plain LLM chat</strong> is mostly “prompt → answer”: the model predicts tokens from the prompt and in-context history. It is <strong>not required</strong> to initiate real-world side effects on its own.</p> |
| 22 | + |
| 23 | +<p>An <strong>agent</strong> wraps the model (and often other parts) so it pursues a <strong>goal</strong>, decomposes work, and when needed <strong>does things</strong>: calls APIs, runs code, searches the web or a document store until the task is done or a step budget is hit.</p> |
| 24 | + |
| 25 | +<p>A useful metaphor: the <strong>LLM is the engine</strong> (reasoning and wording); the <strong>agent is the driver</strong> choosing route, when to stop, and which levers (tools) to pull. An engine without a driving scenario does not take you anywhere by itself.</p> |
| 26 | + |
| 27 | +</div> |
| 28 | +</details> |
| 29 | + |
| 30 | +### What makes an agent {.toc-heading-only} |
| 31 | + |
| 32 | +<details class="post-accordion"> |
| 33 | +<summary style="cursor: pointer; font-weight: 600;">What makes an agent</summary> |
| 34 | +<div style="margin-top: 0.75em;"> |
| 35 | + |
| 36 | +<p>People often sketch three pillars—an “anatomy of an agent” diagram.</p> |
| 37 | + |
| 38 | +<p><strong>Planning.</strong> Breaking a large task into subgoals and steps. This overlaps with <strong>chain-of-thought</strong> style reasoning, explicit plans, and <strong>self-reflection</strong> (re-read a draft, check constraints, adjust).</p> |
| 39 | + |
| 40 | +<p><strong>Memory.</strong></p> |
| 41 | +<ul> |
| 42 | +<li><strong>Short-term</strong> — the live context window: recent turns and intermediate notes in one session.</li> |
| 43 | +<li><strong>Long-term</strong> — stored facts, docs, prior sessions; often implemented with <strong>vector stores and RAG</strong> so relevant snippets are retrieved into the prompt (see the <a class="link" href="/en/p/ai-basics-rag/">RAG article</a>).</li> |
| 44 | +</ul> |
| 45 | + |
| 46 | +<p><strong>Tools.</strong> Formalized actions the model can request: HTTP APIs, code execution (e.g. Python), web search, filesystem hooks, calendar or DB queries. Without tools the “agent” remains text-only with no outward levers.</p> |
| 47 | + |
| 48 | +</div> |
| 49 | +</details> |
| 50 | + |
| 51 | +### How to think about agents {.toc-heading-only} |
| 52 | + |
| 53 | +<details class="post-accordion"> |
| 54 | +<summary style="cursor: pointer; font-weight: 600;">How to think about agents</summary> |
| 55 | +<div style="margin-top: 0.75em;"> |
| 56 | + |
| 57 | +<p>The mindset shift is closer to delegating to a junior teammate than to typing a single search query.</p> |
| 58 | + |
| 59 | +<ul> |
| 60 | +<li><strong>Delegate, don’t only prompt.</strong> State the goal, success criteria, and allowed sources/tools the way you would brief a colleague.</li> |
| 61 | +<li><strong>Set boundaries and a role.</strong> A clear persona (“research assistant”, “no payment actions”) and guardrails reduce scope creep and unsafe surprises.</li> |
| 62 | +<li><strong>Expect iteration.</strong> Agents misstep and dead-end; a normal pattern is try → observe → replan. That drives step limits, logging, and human oversight.</li> |
| 63 | +<li><strong>Provide resources.</strong> If facts live in docs, wire search or RAG; if numbers matter, supply a runtime. Without the right levers the model falls back to guessing.</li> |
| 64 | +</ul> |
| 65 | + |
| 66 | +</div> |
| 67 | +</details> |
| 68 | + |
| 69 | +### ReAct: reason and act {.toc-heading-only} |
| 70 | + |
| 71 | +<details class="post-accordion"> |
| 72 | +<summary style="cursor: pointer; font-weight: 600;">ReAct: reason and act</summary> |
| 73 | +<div style="margin-top: 0.75em;"> |
| 74 | + |
| 75 | +<p><strong>ReAct</strong> (*Reasoning and Acting*) is a common loop drawn as thought → action → observation.</p> |
| 76 | + |
| 77 | +<ul> |
| 78 | +<li><strong>Thought:</strong> what do I know, what is missing, what is the next sensible move?</li> |
| 79 | +<li><strong>Action:</strong> invoke a specific tool with arguments (search, API, code, …).</li> |
| 80 | +<li><strong>Observation:</strong> the tool’s raw result is fed back into context.</li> |
| 81 | +<li>Then another <strong>Thought</strong>—adjust the plan or finish with a final user-facing answer.</li> |
| 82 | +</ul> |
| 83 | + |
| 84 | +<p>This alternates <strong>natural-language reasoning</strong> with <strong>grounded steps</strong> instead of hallucinating when fresh data or computation is required.</p> |
| 85 | + |
| 86 | +</div> |
| 87 | +</details> |
| 88 | + |
| 89 | +### Multi-agent systems {.toc-heading-only} |
| 90 | + |
| 91 | +<details class="post-accordion"> |
| 92 | +<summary style="cursor: pointer; font-weight: 600;">Multi-agent systems</summary> |
| 93 | +<div style="margin-top: 0.75em;"> |
| 94 | + |
| 95 | +<p>Hard problems are sometimes split across <strong>several agents</strong> with different roles—a “manager and workers” picture.</p> |
| 96 | + |
| 97 | +<ul> |
| 98 | +<li>A <strong>coordinator</strong> assigns subtasks, merges outputs, keeps the end result coherent.</li> |
| 99 | +<li><strong>Specialists</strong> might focus on coding, testing, literature search, or report formatting.</li> |
| 100 | +</ul> |
| 101 | + |
| 102 | +<p>Upsides: modularity and parallelism. Downsides: harder debugging, model-call cost, and risk of diverging context between agents. For prototypes, spell out the handoff protocol—who passes what, in what schema.</p> |
| 103 | + |
| 104 | +</div> |
| 105 | +</details> |
| 106 | + |
| 107 | +### Who uses them {.toc-heading-only} |
| 108 | + |
| 109 | +<details class="post-accordion"> |
| 110 | +<summary style="cursor: pointer; font-weight: 600;">Who uses them</summary> |
| 111 | +<div style="margin-top: 0.75em;"> |
| 112 | + |
| 113 | +<p>Agentic setups appear when a single chat reply is not enough—you need an <strong>action chain</strong> tied to tools or corpora.</p> |
| 114 | + |
| 115 | +<ul> |
| 116 | +<li><strong>Researchers.</strong> Scanning many PDFs, summaries, checking phrasing against sources—often with RAG and search.</li> |
| 117 | +<li><strong>Students and academics.</strong> Source discovery and draft surveys with explicit citations—never replacing fact-checking.</li> |
| 118 | +<li><strong>Developers.</strong> Multi-step debugging, refactors, automating ticket/CI/docs flows—with code review and caution.</li> |
| 119 | +<li><strong>Enterprises.</strong> Internal assistants over CRM, wikis, APIs—with strict access policy, action audit trails, and human gates on critical operations.</li> |
| 120 | +</ul> |
| 121 | + |
| 122 | +</div> |
| 123 | +</details> |
0 commit comments