What is an Agent?

An agent is a software entity that perceives context, runs on a schedule or trigger, calls tools, and reasons with a model — all without you having to babysit it.

Agents, chatbots, and workflows

These three are often confused. They are not the same thing.

Workflow

A fixed sequence of steps you design once and run repeatedly. Deterministic — same input, same path. You trigger it; it executes.

Chatbot

Stateless Q&A. Starts fresh every conversation. No memory, no autonomy, no ability to act on the world without your direct instruction.

Agent →

Goal-directed, autonomous, context-aware, and persistent. Decides what to do, when, and in what order — using workflows as its tools.

Agents vs. workflows

Agents use workflows as building blocks. A workflow is just another tool the agent can call.

	Workflow	Agent
Initiated by	You, manually	Heartbeat or trigger
Memory	Stateless	KG + file memory
Adapts	No — fixed steps	Yes — model decides next action
Best for	Defined, repeatable tasks	Open-ended, recurring goals

The four building blocks

Every AgentLed agent is composed of the same four parts. Select any dimension to explore it.

Context

Knowledge Graph + Files

An agent is only as smart as what it knows. In AgentLed, context lives in two places.

The Knowledge Graph (KG) is a persistent, structured memory — entities like companies, contacts, and deals with typed relationships. The agent queries and writes to the KG as it works, so every run starts with accumulated knowledge, not a blank slate.

Markdown files (SOUL.md, MEMORY.md, IDENTITY.md) hold the agent's role definition, persona, and long-term notes. Human-readable, git-friendly, and directly editable.

Knowledge GraphSOUL.mdMEMORY.mdSession state

Real example — Outbound Lead Agent

The agent runs a continuous pipeline: find leads, qualify them, prepare outreach — without anyone managing the queue.

01 —

Daily signal search

Heartbeat: every morning at 7am

The agent wakes up and runs searches across LinkedIn, Crunchbase, and Google News based on your ICP criteria — funding announcements, hiring spikes, new job postings that signal intent, leadership changes. Every matching company gets written to a new_leads list in the KG with the raw signal attached: what was found, where, and when. Nothing is in anyone's inbox yet. The list just grows.

02 —

Qualification run

Workflow: triggered every 2 days

A workflow fires against everything in new_leads that hasn't been qualified yet. For each company it pulls firmographic data, scores against your ICP definition in SOUL.md, checks the KG for existing relationship history, and marks each lead as qualified, low-priority, or disqualified with the reasoning written back to the KG. For qualified leads it drafts a personalised first-touch email using the signal as the hook.

03 —

Monitoring and queue management

Heartbeat: every 6 hours

The agent checks pipeline state: are workflows completing, are any leads stuck, has a qualified lead gone cold? It triggers follow-up workflows, escalates stalled leads to Slack, or pauses outreach if a contact has replied. It also updates sent_outreach and awaiting_reply lists in the KG so the next qualification run doesn't double-touch anyone.

Outbound Lead Agent · Running

Context

KG with new_leads, qualified_leads, sent_outreach, and awaiting_reply lists. SOUL.md defines ICP criteria and outreach tone. MEMORY.md tracks which signals have already been processed.

Heartbeat

7am daily — runs signal search, writes new leads to KG. Every 6 hours — monitors workflow state, manages queue, escalates stalled leads to Slack.

Tools

LinkedIn search, Crunchbase API, web search, KG read/write, email draft, Slack send, workflow trigger.

Model

Claude — scores ICP fit, decides qualify vs. skip, writes personalised email copy using the signal as context, decides when to escalate vs. wait.

Outcome: SDRs start each day with a qualified, prioritised list and draft emails already written — instead of spending 3 hours on research and 2 hours on copy.