Agents · 2023– · models that act
Agent Architecture
A model embedded in a loop: it observes state, plans, calls tools, reads results and revises — with memory and supervision as first-class components.
ModLensAgents
Interactive Diagram
Focus the lens
Click any component to read what it does. Signal direction follows the edges.
inputGoal. A task with completion criteria — not a question.
Core idea
Intelligence plus an action space. The model's job is no longer to answer but to complete: decompose, act, observe, adapt — until the task is done.
Why it exists
Single-shot answers cannot complete multi-system workflows; agency closes the loop between reasoning and action.
Data Flow
What moves through the system
- 01Goal and context enter the planner.
- 02The model proposes a step and selects tools.
- 03Tools execute; observations return to context.
- 04Memory persists state across steps; a supervisor gates risky actions.
Strengths
- + End-to-end workflow completion
- + Composable tool surfaces
- + Progressive autonomy with supervision
Limitations
- − Error compounding across steps
- − Cost and latency per task
- − Evaluation is genuinely hard
Applications
- · Workflow automation
- · Research assistants
- · Operations copilots