Deep Agents

Most AI agents in use today can be described as shallow agents: they handle manageable tasks in a few steps and within a single context window. For simple automation this is sufficient. For complex, long-horizon tasks such as deep research, multi-stage software development, or scientific data analysis, however, they run into structural limits.

Deep agents refer to a new generation of architectures that overcome these limits through strategic planning, persistent memory, and intelligent task delegation.

Why Shallow Agents Fail

A single agent with a long context typically fails for three reasons:

• Context growth: With each step the context expands and eventually exceeds the model’s processing capacity.
• Forgotten steps: On very long tasks, the model loses track of early instructions or overlooks completed sub-tasks.
• Mixed responsibilities: An agent that must simultaneously plan, search, write, and validate makes errors in each of these areas, and those errors accumulate over time.

The Five Characteristics of Deep Agents

1. Structured Planning

Rather than reacting spontaneously within a context window, deep agents maintain a structured task plan, a living to-do list that is updated, extended, and corrected during execution. The plan is stored externally (for example in a file or database) and is therefore independent of the volatile context.

The result: the agent does not lose sight of long-horizon goals and can resume at the right point after interruptions or errors.

2. Orchestrator-Subagent Architecture

A central orchestrator agent handles planning and coordination. It delegates specialised sub-tasks to subagents, each with its own focused context:

    flowchart TB
    O["Orchestrator<br/>(planning and coordination)"]

    O --> R["Research Agent"]
    O --> A["Analysis Agent"]
    O --> V["Verification Agent"]
  

Separating responsibilities not only improves reliability but also enables the targeted use of different models. A powerful but more expensive model handles planning, while lighter models execute repetitive sub-tasks.

3. Persistent Memory and Agentic Search

Deep agents do not rely solely on conversation history. Intermediate results are stored in external memory, for example files, vector databases, or structured notes, and retrieved as needed. This allows relevant information to be reloaded into context without filling it entirely.

Advanced approaches combine semantic search (vector-based, content-similar documents) with structured queries (keyword-based, exact matches) into hybrid retrieval strategies.

4. Context Engineering

The most important lever for the reliability of a deep agent is not the model itself but the quality of its instructions. This discipline is called context engineering and covers the systematic design of system prompts, tool definitions, output formats, and error-handling rules.

Instructions that are too brief, too vague, or contradictory are the most common cause of unreliable agent behaviour. Details are covered in the Context Engineering section.

5. Verification

Verification is the most frequently underestimated component of an agent system. Even powerful models hallucinate, are susceptible to sycophantic behaviour (confirming what the user wants to hear), and can be misled by manipulated tool outputs (prompt injection).

Verification can be implemented in two ways:

• Automated: One LLM evaluates the output of another agent against defined criteria (LLM-as-a-Judge).
• Human: A subject expert reviews critical intermediate results before the next step.

In academic contexts, a combination is recommended: automated plausibility checks for routine operations, human review for professionally relevant or safety-critical outputs.

Case Study: Academic Research System

A deep agent for systematic literature research could be structured as follows: