⑤ 90-Day Plan
⑤ 90-Day Plan
Ninety days is enough to go from “I have read about agent security” to “here are three systems I built, here is how I broke each one, and here is the evidence.” It is not enough to become an expert — and pretending otherwise is how roadmaps produce candidates who interview badly.
The plan below optimizes for a single outcome: at day 90 you have public artifacts that demonstrate you can build an agent, break it, and document the break in the language of a recognized framework.
The most common way this plan fails is spending all three months building and none of it breaking. If you are behind schedule, cut scope from the build, never from the red team work in month two or the documentation in month three. A small agent with a thorough attack write-up beats an ambitious agent with a README.
The three months
- Month 1 — Foundations — Python, Git, Docker, and an AI coding assistant, culminating in the Log Triage Agent: a single agent that reads local logs and summarizes anomalies. Learning checkpoint: self-assess your understanding of at least five OWASP LLM Top 10 entries.
- Month 2 — Orchestration & MCP — multi-agent orchestration, APIs, cloud security, and the Model Context Protocol, culminating in the Multi-Agent SOC: a three-agent pipeline for alert investigation, ticket creation, and report writing, with tools exposed through an MCP server. Then a red team mini-lab where you attack your own agents and log every attempt.
- Month 3 — Go Public — publish three to five projects on GitHub with architecture documentation and a demo recording.
- Required Artifacts — the four deliverables that separate a portfolio from a project list: a STRIDE threat model, a trust boundary and human-approval design, an audit log sample with OpenTelemetry spans, and a “how this agent fails” section in every README.
Python, Git, Docker, AI assistants → Log Triage Agent
Multi-agent pipelines, MCP servers, and red teaming your own work
Publishing, architecture docs, and demo recordings
Threat model, trust boundaries, audit trail, failure analysis
Weekly cadence
| Weeks | Build | Study | Output |
|---|---|---|---|
| 1–2 | Environment, Python and Git fluency, first scripted automation | Layer 2 engineering gaps | A repo with tests and CI |
| 3–4 | Log Triage Agent | OWASP LLM Top 10 | Agent repo + self-assessment notes |
| 5–6 | Multi-Agent SOC pipeline | Orchestration patterns, cloud IAM | Three-agent pipeline running locally |
| 7–8 | MCP server exposing your tools | MCP spec, tool permission design | MCP server + tool permission matrix |
| 9–10 | Red team mini-lab against your own agents | Agentic threat catalogue, ATLAS | Attack log, findings table, audit traces |
| 11–12 | Documentation, diagrams, demo video, publication | NIST AI RMF, EU AI Act basics | Public portfolio + risk register |
The runnable versions of every project referenced here live in Labs.
Day 91 is where most roadmaps end and most careers stall. Before you finish month three, read the continuous learning loop and schedule the weekly and monthly habits — it is far easier to start them while momentum is high than to restart later.