// learn
Agentic engineering, explained simply.
Clear explanations of the ideas behind AI-driven software delivery, written by people who build it every day. No hype, just how it works and what to do about it.
// start here · featured
Why workflow-based execution matters in agentic engineering
Engineering already runs on process. Why breaking work into checked, resumable steps is the right foundation for AI agents.
read · 7 min →Articles
- workflows7 minWhy workflow-based execution matters in agentic engineeringEngineering already runs on process. Why step-by-step execution beats one agent improvising end to end.
- costcoming soonHow to calculate ROI for AI in engineeringA practical method: what to count as cost, what to count as gain, and how to measure it.
- futurecoming soonWhen applications talk to agentsSoftware is becoming agent-callable. What that changes for how teams build.
- foundationscoming soonHarness, orchestrator, platform: what's the difference?Three words used interchangeably, for three different layers. Getting them straight.
- governancecoming soonHITL vs HOTL vs HOOTL: choosing the right human roleHuman in, on or out of the loop. How to decide, step by step, based on risk.
- workflowscoming soonLoop engineering: designing the loop, not the promptThe emerging skill behind reliable agents: plan, act, check, correct.
- costcoming soonModel routing: the right model for each taskWhy the most expensive model on every call is the most common waste in AI engineering.
// no articles on this topic yet
// glossary
The words, in one line each.
Agents & architecture8
| Agent | An AI system that can plan and take actions, such as calling tools, to complete a task. |
|---|---|
| Harness | The runtime that drives an agent's loop: it calls the model, runs tools and limits what the agent can touch. Claude Code and Codex are examples. |
| Orchestrator | The layer that coordinates multiple steps or agents and decides what runs next. |
| Workflow | A defined sequence of steps, with checks and approvals, that turns an input into an output. article → |
| Stateless workflow | A workflow that stores each step's output, so it can pause, resume or retry without holding everything in memory. article → |
| Loop engineering | Designing the cycle an agent follows (plan, act, check, correct) rather than only writing prompts. |
| MCP | Model Context Protocol. An open standard for connecting AI applications to tools and data sources. |
| Agent-to-agent protocol | A standard that lets agents from different systems discover each other and exchange tasks. |
Human oversight & quality6
| HITL | Human in the loop. A person approves or corrects the work before it continues. |
|---|---|
| HOTL | Human on the loop. The work runs on its own while a person monitors and can step in. |
| HOOTL | Human out of the loop. The work runs fully automated within agreed limits. |
| Critic / evals | Automated checks that score AI output against a rubric before it moves on or reaches a person. article → |
| Scorecard | The criteria a critic uses to judge output for a specific workflow. |
| Definition of Ready / Done | Agreed criteria for when a story can start, and when work counts as finished. |
Models, memory & cost7
| Model routing | Sending each task to the model that best fits its quality and cost needs. article → |
|---|---|
| Model Rail | Sage's single governed route for every model call, used to apply model policy and attribute cost. |
| Token | The unit of text a model reads and writes. Usage and cost are measured in tokens. |
| Context window | The amount of text a model can consider at once. |
| Short-term memory | Context from recent runs, used to keep continuity across a few tasks. |
| Long-term memory | Confirmed learnings kept over time and reused across many runs. |
| Codebase knowledge store | A structured map of a codebase (services, components, conventions) that agents can draw on. |
Governance & security5
| Tool Gateway | Sage's single governed route for every action on external systems such as repos and tickets. |
|---|---|
| Prompt versioning | Managing prompts like code, with versions, owners, review and rollback. |
| Skills | Reusable, packaged instructions an agent can load for a specific kind of task. |
| Prompt injection | An attack where text in the input, such as a ticket or PR, tries to override an agent's instructions. |
| SBOM | Software bill of materials. A list of every component and dependency in a piece of software. |
// no matching terms
26 terms
// put it into practice