Zola 2026-05-12T00:00:00+00:00 https://hanlho.com/tags/tool-calling/atom.xml 2026-05-12T00:00:00+00:00 2026-05-12T00:00:00+00:00 Unknown https://hanlho.com/p/when-to-use-expert-tool-calling-llms/ <p>Often, large language models (LLMs) are evaluated in certain categories, for example, coding, writing, or tool-calling. I wasn't sure what tool-calling meant, so I decided to do some AI-assisted research. I used <a href="https://chat.mistral.ai/chat">LeChat</a> and GPT-5.4, going back and forth between them, checking references, to get a better understanding.</p> <p>My main desired outcome was an infographic that is usable, and I think it does a good high-level job of explaining what tool-calling is and when you should use it. That is why I decided to share it here as well.</p> <p>To find LLMs to use for this category you could check <a href="https://llm-stats.com">LLM Stats</a>.</p> <p>To be clear and explicit: <em>the below is generated by an LLM but driven and reviewed by me</em>. I also added a text version that is an almost exact copy after the graph.</p> <h1 id="llm-tool-calling">LLM Tool-Calling</h1> <p><img src="/img/info-when-to-use-expert-tool-calling-llms.png" alt="" /></p> <p><strong>Here is <em>the same content</em> in text:</strong></p> <h1 id="when-to-use-expert-tool-calling-llms">When to Use Expert Tool-Calling LLMs</h1> <p><em>A quick guide to deciding where tool-calling models create the most value.</em></p> <hr /> <h2 id="1-final-rule">1. Final Rule</h2> <blockquote> <p><strong>Use a tool-calling LLM when the task is not just <em>“answer this”</em> but <em>“go get information, act on it, inspect what happened, and decide the next step.”</em></strong></p> </blockquote> <hr /> <h2 id="2-decision-heuristic">2. Decision Heuristic</h2> <p>Use a tool-calling model when the task requires <strong>at least 2</strong> of the following:</p> <blockquote> <p><strong>Rule of thumb:</strong><br /> <strong>2+ checks = strong fit</strong></p> </blockquote> <table><thead><tr><th style="text-align: right">Check</th><th>Heuristic</th><th>What it means</th></tr></thead><tbody> <tr><td style="text-align: right">✅</td><td><strong>External state access</strong></td><td>Needs live or private data from code, databases, logs, tickets, APIs, calendars, or documents.</td></tr> <tr><td style="text-align: right">✅</td><td><strong>Iterative execution</strong></td><td>Requires action → inspect result → refine → rerun.</td></tr> <tr><td style="text-align: right">✅</td><td><strong>Structured action</strong></td><td>Needs API calls, SQL queries, JSON arguments, file edits, or workflow steps.</td></tr> <tr><td style="text-align: right">✅</td><td><strong>Cross-system synthesis</strong></td><td>Combines information or actions across multiple tools or systems.</td></tr> <tr><td style="text-align: right">✅</td><td><strong>Verifiable output</strong></td><td>Produces something checkable: tests, tickets, records, charts, or citations.</td></tr> </tbody></table> <hr /> <h2 id="3-top-3-high-differentiation-use-cases">3. Top 3 High-Differentiation Use Cases</h2> <h3 id="1-multi-system-workflow-automation">1. Multi-System Workflow Automation</h3> <p><strong>Best when</strong> work spans multiple tools and requires conditional logic, structured API calls, and stateful execution.</p> <p><strong>Examples</strong></p> <ul> <li>GitHub issue → Jira ticket → Slack notification</li> <li>Support ticket → order lookup → refund check → escalation</li> <li>CRM update → email draft → calendar follow-up</li> </ul> <hr /> <h3 id="2-codebase-exploration-debugging-maintenance">2. Codebase Exploration, Debugging &amp; Maintenance</h3> <p><strong>Best when</strong> the model must inspect code, follow references, run tests, read logs, and compare implementation to specs.</p> <p><strong>Examples</strong></p> <ul> <li>Trace authentication flow</li> <li>Investigate failing tests</li> <li>Find affected files for a feature change</li> <li>Draft a patch with file references</li> </ul> <hr /> <h3 id="3-structured-data-analysis-operational-analytics">3. Structured Data Analysis &amp; Operational Analytics</h3> <p><strong>Best when</strong> the model needs to query, transform, analyze, or visualize live data.</p> <p><strong>Examples</strong></p> <ul> <li>Generate and refine SQL</li> <li>Analyze churn or revenue by segment</li> <li>Run Python anomaly detection</li> <li>Create a chart and explain outliers</li> </ul> <hr /> <h2 id="4-do-not-use-a-tool-calling-llm-primarily-for">4. Do NOT Use a Tool-Calling LLM Primarily For</h2> <table><thead><tr><th>Avoid</th><th>Reason</th></tr></thead><tbody> <tr><td><strong>Pure writing or ideation</strong></td><td>Better handled by a writing-optimized or general-purpose model.</td></tr> <tr><td><strong>Summarizing text already provided</strong></td><td>No external data access or tool use is required.</td></tr> <tr><td><strong>Static Q&amp;A without external data</strong></td><td>A general reasoning model is usually sufficient.</td></tr> <tr><td><strong>High-risk production changes without approval</strong></td><td>Requires human oversight, review, and explicit approval.</td></tr> <tr><td><strong>Legal, medical, or financial decisions without expert review</strong></td><td>Tool use can support retrieval and analysis, but expert validation is required.</td></tr> </tbody></table> <hr /> <h2 id="compact-decision-rule">Compact Decision Rule</h2> <p>Use an expert tool-calling LLM when the task involves:</p> <ol> <li><strong>Getting information</strong> from live or private systems.</li> <li><strong>Taking structured action</strong> through tools, APIs, queries, or file operations.</li> <li><strong>Inspecting results</strong> and adapting the next step.</li> <li><strong>Producing verifiable output</strong> such as tests, tickets, records, charts, or citations.</li> </ol> <p>If the task is only to write, summarize provided text, or answer from static context, a tool-calling LLM is usually unnecessary.</p>