Agent Loop

The Agent Core Loop 🔄

Gideon doesn’t just follow a script. It operates using an autonomous reasoning loop based on the ReAct (Reasoning and Acting) framework.

How it Works

The loop follows four distinct stages in every iteration:

1. Planning (Reason)

Gideon reviews the current goal and any accumulated context. It decides on the next logical step and explains its reasoning in the scratchpad.

2. Execution (Act)

Gideon selects the most appropriate tool for the task (e.g., cve_search, ioc_lookup) and executes it with structured parameters.

3. Verification (Observe)

The result of the tool execution is fed back into the loop. Gideon analyzes the output:

• Is the data sufficient?
• Does it contradict previous findings?
• What is the confidence score?

4. Refinement (Reflect)

Gideon asks itself: “What would change my mind?” It tracks assumptions and seeks cross-source corroboration before finalizing an answer.

---

Visualizing the Loop

graph TD
    User[User Query] --> Planner[Task Planner]
    Planner --> Reasoning[Reasoning Step]
    Reasoning --> Action[Action/Tool Call]
    Action --> Observation[Observation/Tool Result]
    Observation --> Corroboration[Cross-Source Validation]
    Corroboration --> Reflection{Goal Met?}
    Reflection -- No --> Reasoning
    Reflection -- Yes --> Final[Final Answer & Report]

Security Guardrails

The loop is wrapped in a safety layer that monitors every input and output:

1. Pre-Execution Check: Rejects any tool calls that appear to target offensive operations.
2. Confidence Scoring: If corroboration fails or confidence is low, Gideon explicitly flags the result as an “Assumption.”
3. Topic Steering: NVIDIA NeMo Guardrails ensure Gideon stays on-topic (Defensive Cybersecurity) and avoids hallucinated instructions.