Agents vs Workflows
Workflows codify how work gets done. Agents decide what work to do, when, and in what order — using workflows as tools.
The Core Distinction
A workflow is a translation of a process into executable steps. It is deterministic: given the same input, it follows the same path. Trigger → steps → output. You design it once, run it repeatedly.
An agent is a reasoning layer on top of workflows. It reads context, decides which workflows to invoke, sequences them, monitors their outputs, evaluates results, and loops until the goal is reached. The agent owns the outcome; workflows are its instruments.
Workflow
- •Codified process — structured, repeatable steps
- •Triggered explicitly (schedule, webhook, manual)
- •Executes a defined path
- •Returns output — does not own outcomes
- •Composable — used as a tool by agents or other workflows
Agent
- •Goal-directed — decides what to run and when
- •Proactive — acts on schedule, signals, or memory triggers
- •Sequences and adapts based on intermediate results
- •Owns outcomes — monitors, evaluates, retries
- •Learns — reads KG context, writes decisions back
Workflows as Tools
In AgentLed, any workflow can be registered as a tool available to an agent. The agent does not execute steps directly — it invokes workflows, receives their outputs, and decides what to do next based on the results.
This separation matters for two reasons:
- 1.Reuse. The same “enrich contact” workflow can be invoked by a prospecting agent, a re-engagement agent, and a data quality agent — without duplicating any logic.
- 2.Observability. Every workflow execution is a discrete, traceable unit. You can inspect inputs, outputs, duration, and errors at the workflow level — regardless of which agent triggered it.
Example: Prospecting Agent
A prospecting agent owns the goal of filling the top of the pipeline with qualified leads. It does not contain any enrichment or outreach logic itself — it orchestrates workflows that do.
Read KG context
Agent queries Knowledge Graph: which companies match ICP? Which have been contacted in the last 30 days? Which had responses?
Invoke: enrich-company workflow
For each new target company, runs the enrichment workflow. Returns firmographic data, tech stack, recent hiring signals.
Invoke: score-lead workflow
Scores each enriched company using the lead scoring workflow. Compact profile from KG included automatically.
Evaluate results
Agent checks scores against threshold. Companies above threshold advance. Below threshold are written back to KG with reason for skip.
Invoke: draft-outreach workflow
For approved leads, runs the outreach draft workflow. Personalizes based on enrichment data and KG history.
Human gate (optional)
Agent routes drafts to human-in-the-loop review if configured. Waits for approval before sending.
Monitor and write outcomes
Agent tracks sends, opens, replies. Writes outcomes back to KG as CONTACTED and OUTCOME edges. Adjusts scoring context for next run.
The agent owns the pipeline goal. Each workflow is a focused, reusable unit. Neither knows about the other — the agent composes them at runtime.
Execution Monitoring
Every agent run creates a timeline of workflow invocations. You can inspect:
| Signal | What it tells you |
|---|---|
| Workflow invocations | Which workflows the agent called, in what order, with what inputs |
| Step-level outputs | Output of each step within a workflow execution — inspectable without re-running |
| Agent reasoning | Why the agent chose a particular workflow or branched in a given direction |
| Errors and retries | Which workflows failed, whether the agent retried, and with what result |
| KG writes | Every node and edge the agent created — auditable after the fact |
Performance Evaluation
Because agents write decisions and outcomes to the Knowledge Graph, performance can be evaluated over time — not just per-run:
- •Scoring accuracy. Compare AI scores to human decisions and downstream outcomes. Identify where the agent is systematically wrong.
- •Workflow utilization. Which workflows does the agent invoke most? Which are rarely used? This surfaces dead weight and missing capabilities.
- •Outcome tracking. Connect agent decisions to real-world results — meetings booked, deals closed, hires made. Measure whether the agent is moving the right metric.
- •Correction rate. How often do humans override agent decisions? High correction rates on specific entity types indicate rubric drift or missing context.
When to Use Each
Use a workflow when
- •The process is well-defined and repeatable
- •You want full control over the execution path
- •The task does not require reasoning about context or goals
- •You are building a reusable capability that agents or other workflows will invoke
Use an agent when
- •You are defining a goal, not a process
- •The right sequence of steps depends on intermediate results
- •You need the system to monitor, retry, and adapt — not just execute once
- •You want compound learning — outcomes feeding back into future decisions
Next Steps
- Create & Configure Agents — Set up an agent and register workflows as its tools
- Agents as Step — Use an agent as a step inside another workflow
- Multi-Agent Orchestrator — Coordinate multiple agents toward a shared goal
- KG Overview — How agents read and write memory across runs
- Scoring & Feedback — The learning loop agents use to improve over time