Skip to content

LOCKEDIN LABS / APPLIED INTELLIGENCE

Move from
AI possibility
to production.

Learn the decisions behind AI that people can use, teams can operate and leaders can stand behind.

Find your learning path

Free interactive field study · no account or API key required

SERVICECONTEXTAGENTDECISIONONE PRODUCTION SYSTEM / FOUR PERSPECTIVES
Architecture study · select a role to inspect its responsibility
Agent engineeringEvaluation & observabilityEnterprise deliveryHuman ownership

01 / CHOOSE YOUR RESPONSIBILITY

Different roles.
A shared production standard.

Follow your own path, then bring your evidence to the people receiving the work.

YOUR PERSPECTIVE / AI engineer

Turn model capability into dependable behavior.

Design agent workflows, retrieval and tools that hold their boundaries when inputs, models and network conditions change.

  • Agent state & recovery
  • MCP tool boundaries
  • Retrieval & evaluation

FIELD NOTE / 01

The model proposes; the application owns authority

A useful agent can choose a lookup or propose a tool call. Your application still authenticates the caller, resolves the permitted tenant and asset, validates the action and binds it to a specific request. A plausible tool argument is input to those checks, not proof of permission. Keep retrieved instructions separate from the controls deciding which sources and actions the caller may use.

For a Model Context Protocol (MCP) integration, diagram the client, server and downstream resource separately. Name the token audience and delegated permissions at each boundary. A token accepted by one resource is not automatically valid for the next. Test the legitimate request as well as the forbidden one; a system that rejects everything has not implemented a useful workflow.

FIELD NOTE / 02

Evaluate the path and the resulting state

A fluent final answer can conceal a wrong source, repeated side effect or missing approval. Keep a small versioned case set with expected intermediate decisions and independently observable outcomes. Run repeated trials when behavior varies. Compare candidates on the same tasks and distinguish grader errors from system failures.

In this studio, a response saying “draft saved” is weaker evidence than a matching stored receipt. If a save times out, inspect that original request before proposing a new operation. Build a test in which the receiver cannot tell whether the write committed; a retry should preserve request identity and follow a demonstrated idempotency contract.

EXECUTABLE EVIDENCE / AI engineer

Distinguish a useful agent from a passing demo

Run the existing lab on your computer. Predict first, keep the actual output, then ask another person to reproduce it.

Python 3.11+
Standard library
Synthetic inputs

01 / PREPARE & RUN LOCALLY

Full exercise & setup

Download every listed file into the same disposable folder. No account, API key or installed package is needed.

python3 --version
python3 fde-agent-evaluation-lab.py run --candidate reference > reference.json
python3 fde-agent-evaluation-lab.py run --candidate refuse-all > refuse-all.json
# The refuse-all candidate intentionally exits 2; retain its report.

Actual temporary SQLite workflow execution with scripted drafts and deliberately broken candidates. No model, network, authentication service or production-readiness verdict.

02 / YOUR INDIVIDUAL RECORD

Will a candidate that refuses every request pass the same contract as a useful reference? Name a useful-work case and a boundary case before running either.

Compare decision, planned, executed, slices and each cases row. Separate a failed observation from an omitted or unrun case; retain both JSON files and their exit statuses.

Have the receiving engineer run the two candidates independently, then challenge the conclusion by omitting one contract case. Keep their own commands and observations.

Notes stay in this page until downloaded. The studio does not run code, inspect your files or verify these self-reports. Keep fictional details and retain the original reports separately.

Continue into the advanced curriculum

Academy lessons · training access required
  1. 01

    Agent engineering with LangChain & LangGraph

    Choose the workflow before the framework

    Specify the allowed work before selecting a framework.

  2. 02

    Build and operate enterprise AI

    Make tools safe when text is hostile

    Test tool authority against hostile input.

  3. 03

    Agent engineering with LangChain & LangGraph

    Evaluate the task, the trajectory and the failure

    Connect trajectory checks to the final outcome.

02 / PUT THE THINKING TO WORK

Would you release it?

One fictional maintenance copilot. Three evidence packets. Explain a decision from your selected role.

SYNTHETIC TRAINING CASE / AI engineer

The persuasive demo

Field Service Copilot v0.8 drafts maintenance plans for fictional asset PUMP-7. It may save a review draft; it may not dispatch work or close a work order. Decide whether this candidate can enter the stated draft-only trial.

These are authored fixture observations, not live telemetry or learner results.
  1. E1

    Task contract

    The receiver requires current manual revision 5 and a human review before any operational action.

  2. E2

    Evaluation packet

    Ordinary in-scope cases ran. Wrong-asset and stale-manual cases were not run.

  3. E3

    Observed output

    The demo draft cites manual revision 4. The required revision is 5.

  4. E4

    Receiving rehearsal

    A recovery document exists; the receiving operator has not tried it.

Your decision for this evidence

Notes stay in this page until you download them. Use fictional details, without client data or credentials.

Bring your own reasoning to the team.

Each person keeps an individual record. Ask a peer to change one fact, challenge the decision and record a revision. Practice does not establish deployment qualification.

03 / KEEP THE SOURCE CLOSE

Built on current
engineering practice.

Primary references checked 2026-11-02. Teaching, scenarios and reference decisions are original LockedIn Labs material. Technical references do not imply partnership or endorsement.

  1. Anthropic · Demystifying evals for AI agents

    Task, trial, grader and outcome distinctions; inspect trajectories as well as final answers.

  2. Model Context Protocol · Security best practices

    Delegation boundaries and token-audience validation; protocol support alone does not authorize a downstream action.

  3. OpenTelemetry · GenAI semantic conventions

    Current primary home for generative AI, agent and MCP signal conventions. Pin the version used by your instrumentation.

Explore the full course catalog ↗Continue with free guides and labs ↗