> For the complete documentation index, see [llms.txt](https://readme.dhee.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://readme.dhee.ai/concepts-we-work-on/agentic-llms.md).

# Agentic LLMs

Large Language Models are powerful at understanding and generating language.\
However, **language alone is not enough** to solve real-world problems.

**Agentic LLMs** extend traditional LLMs by giving them the ability to **decide, act, observe outcomes, and adapt** — much like a human operator executing tasks step by step.

Instead of responding once and stopping, an agent **reasons over multiple steps**, uses tools, and works toward a goal.

<figure><img src="/files/12VjflnCfVaSd8uzxb6r" alt=""><figcaption></figcaption></figure>

### Why Agentic LLMs Exist

A normal LLM answers questions.

An agentic LLM answers questions **and then does something about it**.

Examples:

* Reading a document → extracting facts → validating them → storing results
* Understanding a user request → choosing tools → calling APIs → verifying outputs
* Planning a multi-step task → executing → correcting mistakes → finishing the goal

Agentic behavior is essential when:

* The task **cannot be completed in one response**
* The system must **interact with external systems**
* The model must **self-correct or re-plan**

### Core Agent Loop

At the heart of an Agentic LLM is a simple but powerful loop:

1. **Observe**\
   Receive user input, system state, or tool outputs.
2. **Reason**\
   Decide what to do next based on goals and context.
3. **Act**\
   Call a tool, query a database, read a document, or ask a follow-up question.
4. **Evaluate**\
   Check whether the action helped achieve the goal.
5. **Repeat or Stop**\
   Continue until the objective is satisfied.

This loop turns a passive model into an **active problem solver**.

### Tools as Extensions of Cognition

In agentic systems, tools are not add-ons — they are **extensions of the model’s capabilities**.

Common tool types:

* Search engines
* Databases
* Code execution
* APIs
* Memory stores
* Document readers

The LLM does not *know* everything.\
Instead, it **knows how to find, verify, and combine information**.

### Planning vs Execution

Agentic LLMs often separate thinking into two layers:

* **Planner**
  * Breaks a goal into sub-tasks
  * Chooses execution order
* **Executor**
  * Performs each step
  * Reports results back to the planner

This separation improves:

* Reliability
* Debuggability
* Control over long-running tasks

### Memory in Agentic Systems

Unlike single-turn chatbots, agents require **memory**.

Types of memory:

* **Short-term**: current task context
* **Working memory**: intermediate results
* **Long-term**: user preferences, learned facts, prior executions

Memory allows agents to:

* Avoid repeating mistakes
* Maintain continuity
* Learn from previous runs

### Failure Is a Graceful

Agentic systems are **designed to fail safely**.

Instead of collapsing on errors, they:

* Detect failures
* Re-evaluate assumptions
* Retry with alternative strategies

This is critical for:

* Automation
* Enterprise workflows
* Mission-critical systems

### Agentic LLMs and Indian Languages

Most large language models are optimized primarily for **English-first interaction**.\
However, real-world conversational systems — especially in India — require **native competence across multiple Indian languages**, dialects, and code-mixed usage.

At Dhee, we work with C**onversational Agentic LLMs trained and adapted for Indian languages**, focusing on:

* Natural, spoken-style conversations
* Language-specific grammar and morphology
* Cultural context and usage patterns
* Multi-turn dialogue and Task following robustness

These models are designed not just to *translate*, but to **converse natively**.

You can explore our open model collection here:\
👉 [https://huggingface.co/collections/dheeyantra/](https://huggingface.co/collections/dheeyantra)