The premise
An LLM stops being a chatbot and becomes an operator: it plans, calls tools and data, observes results, and iterates toward an outcome.
Agentic engineering is the shift from AI that responds to prompts to AI that pursues goals — systems that reason, use tools, write and run code, check their own work, and complete multi-step tasks with a human in the loop. This page is itself an example: a live agent is embedded here. Ask it questions — or tell it to restyle this page, generate imagery, even replace the video with one it makes on the spot.
An LLM stops being a chatbot and becomes an operator: it plans, calls tools and data, observes results, and iterates toward an outcome.
Tool calling, reasoning, evals, and protocols like MCP matured together in 2023–2025 — turning a research idea into deployable enterprise software.
New capability is also a new trust boundary. Governance, least-privilege access, and human oversight separate leverage from exposure.
What "agentic" actually means
In a traditional workflow, developers wire the steps in code. In an agentic system, the model decides the steps: it chooses which tools to call, reads what comes back, corrects course, and keeps going until the task is done — or hits a guardrail you set.
Anthropic's widely-used framing draws the line cleanly: workflows follow predefined paths; agents dynamically direct their own. The discipline of building, constraining, and evaluating those systems reliably is agentic engineering.
"The most successful implementations use simple, composable patterns rather than complex frameworks."— Anthropic, Building Effective Agents, Dec 2024
Framed for a healthcare security leader
The durable signal is benchmarked task completion (SWE-bench, τ-bench), not demos. Evaluate agents on verified work.
MCP and tool-calling are where an agent touches data and systems — exactly where least-privilege and auditing belong.
The mature pattern is constrained autonomy: agents propose and act within scopes; people approve consequential steps.
Value shows up when agents run inside a governed "mesh" — identity, permissions, observability — not as ungoverned point tools.
The centerpiece · Watch & learn
Pick a talk — it plays right here. Every video was verified as real and embeddable. (Tip: ask the agent to "replace this video with an AI-generated one.")
Select a talk
Behind the build
Agentic engineering isn't just the subject of this page — it's how the page was made. From one plain-English request, an AI agent researched the field and verified every source, wrote and deployed this site, provisioned its own cloud hosting and a least-privilege deploy credential, and built the live assistant you're using now. The full prompt-by-prompt story is a genuinely good read.
Capability vs. hype
The honest read for a decision-maker: coding and tool-use agents have moved from novelty to measurable, fast-improving capability — while reliability, oversight, and governance remain the gating factors for enterprise deployment.
Where it lands in a health system
Prior authorization, documentation, revenue cycle, IT operations, and patient access are exactly the tool-heavy, multi-step processes agents are built for — and exactly where PHI raises the governance bar. Illustrative use cases, not endorsements of specific tools.
The CISO's checklist
If agents can act, the questions a security leader asks shift from "what did the model say?" to "what is the model allowed to do, as whom, with what data, and who approved it?" A practical frame:
The real arc
Every milestone is dated to its first publication and links to a primary source.
Who is building this
Read the primary sources
Papers, standards, and enterprise reports — each link checked to resolve.
In their own words
Real, attributed quotes — speaker, venue, and date — each linked to where it can be verified.
Generated live, by the agent
Anything the embedded agent generates — images (Imagen 4) or video (Veo) — appears here. Ask it: "generate an AI image of a secure healthcare data mesh," or "make an AI video for the hero."
Traceability
Researched before written. Every webinar, paper, and quote traced to a primary source and link-checked.