Agentic Engineering — Field Guide

One page for co-developers working with coding agents.

theme D
01

Your cockpit — what you steer, what it sees, what it can do

one agent session, three channels

How you steeryour controls

Agent harness

The app around the model — VS Code, CLI, cloud agent — that hands it files, tools and UI.

pickSame models and capabilities, different interface · VS Code for built-in visual diffs · CLI for scripting and long unattended runs.

Agent modes

Ask mode explains without touching code · Edit mode changes one file fast · Agent mode plans, edits across files, runs commands and self-corrects.

hintUse agent mode for most tasks, ask mode when you just need an answer.

What the agent seesits context

Context window

The model's working memory: every instruction, file and tool result it can see this turn.

hintSmaller context is cheaper, faster and usually sharper.

System prompt

The provider's baseline instructions — invisible to you, sent with every single request.

hintYou can't edit it; your files layer on top.

Instruction file(s)

Persistent project rules the agent re-reads every session, so you don't repeat yourself.

filesAGENTS.mdcopilot-instructions.mdCLAUDE.mdSame concept, different name — check which one your harness/agent reads.

Skills

Packaged expertise — instructions plus assets. Only the short description sits in context until the task needs it.

try/skillse.g. PowerPoint skill by Anthropic to create and modify pptx files

Specific files

The files you attach yourself — the fastest way to point the agent at the right code.

try#file:Service.cstype # in Copilot chat to attach a file, folder or selection.

Plugins

Bundles that extend the harness — skills, tools and workflows shipped in one install.

hintEach plugin preloads instructions into your window. Install sparingly.

What it can doits reach

Tools

The concrete actions it can take: read, edit, run a command, search, fetch a URL.

hintAgents really like ripgrep (rg), an efficient modern CLI text search tool — install it on your system and provide it in the shell.

MCP

A standard way to connect the agent to other systems — package feeds, ticket trackers, databases.

try/mcpThe NuGet MCP allows your agent to autonomously check for latest and compatible package versions.
02

Ground rules

non-negotiable
  • Less vibe coding, more agentic engineering. Define hard constraints up front, engineer the prompt and context, review the diff, keep it traceable.
  • Keep instruction files small. Below 150 lines of key instructions are usually enough. Anything you won't maintain drifts out of sync and quietly misleads the agent.
  • Feed learnings back to the knowledge base. So the company chatbot — and the next colleague — already know the answer.
  • Don't commit walls of AI documentation. It's verbose, duplicated, and nobody keeps it current.
  • Skip AI where the IDE is exact. Rename symbol, extract method, reformat — deterministic refactorings beat a guess.
  • Keep static & dynamic analysis running. The agent writes the code; your toolchain still has to prove it.
03

Setup & workflow

the loop
  • Choose your harness deliberately. It's an interface choice, not a power ceiling — both run long, heavy tasks. VS Code gives visual diffing with no extra tooling; the CLI scripts, pipes and runs unattended.
  • Brainstorm before you build. Use the AI as a sparring partner to challenge your plan and refine your approach. If possible, use the company ChatBot for this to save credits on your agent plan.
  • Let it write the artefacts of that talk. A concise instruction file, plus a precise implementation prompt for the big task.
  • Expensive model to plan, cheap model to implement. Read the plan at least coarsely before approving — and skip planning entirely on simple, single-file work.
  • Review what lands — don't rubber-stamp. Skim the diff, double-check the key parts: you merge it, you own it.
  • Hand it the tests. AI drafts them fast and covers cases you'd skip — check the assertions encode the requirement, not just current behaviour.
04

Context & tokenomics

Your window before you type a wordfree ⟶
system prompt instruction files mcp · tools · plugins
  1. 1
    Start a fresh session per task. Long threads rot: as the window fills, output quality drops — hardest on small models.
  2. 2
    Write one dense instruction file. Mature codebase or greenfield prototype, target architecture, stack and banned dependencies, known edge cases and workarounds.
  3. 3
    Trim your surface. Every MCP server, tool and plugin eats the window before you type your first word.
  4. 4
    Right-size the model. Small models like GPT Luna are already very capable and a fraction of the price.
  5. 5
    Watch the meter. Your harness shows session cost in AIC — 100 AIC = CHF 1.
  6. 6
    Never solve the same problem twice. After a hard fix, have the agent write the hurdle and the solution into the README or knowledge base.
05

Security

the blast radius
SENSITIVE DATA UNTRUSTED INPUT OUTBOUND ACTIONS

The red overlap is the exploitable case: reads your secrets + reads an untrusted ticket + can act outward → one poisoned issue is enough. Break the overlap and the attack dies.

  • Never paste secrets into a prompt. If it already happened, rotate the affected credentials today.
  • Keep secrets out of the repo. The agent reads everything you gave it access to — including that old config.
  • Treat fetched content as hostile. Web pages, issues and docs the AI pulls in can carry injected instructions.
  • Minimise permissions. Read what you're approving before you click allow — especially shell commands and network calls.
Prompt injection
ruleFetched ≠ trusted.

Untrusted content the agent reads — web pages, issues, third-party READMEs — can smuggle instructions into the context and be mistaken for your orders.

06

Do / don't

Do

  • Plan first on anything multi-file
  • Keep diffs small and reviewable
  • Select context deliberately
  • Run lint and tests yourself
  • Document what was hard to solve

Never

  • Merge code you don't understand
  • Paste secrets or customer data into a chat
  • Swallow a 500-line diff in one go
  • Argue with a lost agent — a fresh session beats a poisoned one
  • Ship AI-written code or docs unread