Skip to content
Explore NVIDIA, Google Cloud, AWS and Azure courses →

Forward-deployed engineer formation

Learn the method. Ship the work.

Eight modules. Production-shaped missions. One standard for evidence that survives outside the classroom.

Open the training demonstration Inspect the qualification bar
Authored modules
08
DEPLOY moves
6
Authored core
14h

Formation is authored. Live cohort delivery activates on cohort launch; qualification issuance activates on preregistration.

LOCKEDIN / FORWARD-DEPLOYED ENGINEERING

Learn the stack.
Own the delivery.

Technical depth meets client judgment. Follow a route through working lessons, guided examples and downloadable labs.

Your starting point
→ Forward-deployed engineer

From shipping code to owning client delivery.

Extend your engineering practice with stateful agents, tool boundaries and the judgment to put a useful system in another team's hands.

  1. 01Claude Code
  2. 02LangChain & LangGraph
  3. 03MCP workflow building
  4. 04Agentic engineering
  5. 05Client delivery

Suggested learning sequence · Start from your role; practise the full delivery loop. Completion alone does not confer field qualification.

01 / TECHNICAL TRACK

Agent-assisted software engineering

Claude Code

Orient in a repository, make a bounded change, review the diff and hand over tested behavior.

WHAT YOU MAKEA reviewed change with reproducible test evidence
Start with 1017 connected lessons
102 · Explore the guided lesson sequence
  1. 01Terminal 101: understand the command before the agent35 min
  2. 02Get oriented before the first edit20 min
  3. 03Write context that a teammate can verify20 min
  4. 04Plan and implement one bounded change35 min
  5. 05Use failures to debug the decision30 min
  6. 06Decide what a tool may do25 min
  7. 07Hand over a change someone can accept25 min

Academy lessons open with a provisioned training seat.

02 / TECHNICAL TRACK

Agents, state and durable workflows

LangChain & LangGraph

Build a tool-calling agent, model explicit state, and recover a human decision across a restart.

WHAT YOU MAKEA checkpointed workflow with an interrupt and recovery path
Start with 1017 connected lessons
102 · Explore the guided lesson sequence
  1. 01Agents, workflows and frameworks: start here25 min
  2. 02Choose the workflow before the framework30 min
  3. 03LangChain: your first agent, explained40 min
  4. 04Build a LangChain agent with bounded tools170 min
  5. 05LangGraph: your first workflow, explained40 min
  6. 06Model state and control flow with LangGraph120 min
  7. 07Recover workflows and bind human decisions60 min

Academy lessons open with a provisioned training seat.

More labs in this track 7

Choose the boundary you want to practise. Each walkthrough includes setup and local commands.

  • Choose the workflow before the framework

    Durable workflow reference

    Pause a draft, restart the process and bind a review decision to its exact source.

    One Python file · No package installation

  • Build a LangChain agent with bounded tools

    MCP tool admission

    Challenge changed tool definitions, forged arguments and current host permissions.

    Node.js 24 · Offline host-policy exercise

  • Model state and control flow with LangGraph

    LangGraph interrupt and resume

    Wire a real StateGraph, observe an interrupt and resume its in-memory state.

    Python 3.11+ · New virtual environment · Pinned LangGraph package

  • Build a LangChain agent with bounded tools

    LangChain tool-calling loop

    Inspect real create_agent orchestration using scripted tool proposals and results.

    Python 3.11+ · New virtual environment · Pinned framework packages

  • Build a LangChain agent with bounded tools

    Structured agent output

    Run real ToolStrategy validation and repair, then distinguish a valid shape from current source support.

    Python 3.11+ · New virtual environment · Pinned framework packages

  • Build a LangChain agent with bounded tools

    MCP stdio compatibility

    Inspect real framed requests and distinguish a tool refusal from a protocol failure.

    Node.js 24 · Owned child process · Pinned 2025-11-25 profile

  • LangGraph: your first workflow, explained

    Your first LangGraph workflow

    Trace two ordinary functions, field updates and a missing-source branch before adding persistence.

    Python 3.11+ · Isolated environment · LangGraph 1.2.16
03 / TECHNICAL TRACK

Model Context Protocol · tools & integration

MCP workflow building

Connect an agent to a discoverable tool without letting model arguments choose its permissions.

WHAT YOU MAKEA tested tool-admission contract with refusal cases
Start with 1015 connected lessons
102 · Explore the guided lesson sequence
  1. 01MCP: from a tool description to a useful result45 min
  2. 02Build a LangChain agent with bounded tools170 min
  3. 03RAG: from documents to model context35 min
  4. 04Embeddings: from a sentence to a search result45 min
  5. 05Build retrieval and memory that respect source boundaries75 min

Academy lessons open with a provisioned training seat.

More labs in this track 5

Choose the boundary you want to practise. Each walkthrough includes setup and local commands.

  • Build retrieval and memory that respect source boundaries

    Scoped retrieval and context

    Compare retrieval quality on eight cases while preserving current access and context budgets.

    Python standard library · Standalone file

  • Build a LangChain agent with bounded tools

    LangChain tool-calling loop

    Inspect real create_agent orchestration using scripted tool proposals and results.

    Python 3.11+ · New virtual environment · Pinned framework packages

  • Build a LangChain agent with bounded tools

    Structured agent output

    Run real ToolStrategy validation and repair, then distinguish a valid shape from current source support.

    Python 3.11+ · New virtual environment · Pinned framework packages

  • Build a LangChain agent with bounded tools

    MCP stdio compatibility

    Inspect real framed requests and distinguish a tool refusal from a protocol failure.

    Node.js 24 · Owned child process · Pinned 2025-11-25 profile

  • MCP: from a tool description to a useful result

    MCP SDK client and server

    Run real SDK discovery in two protocol profiles and separate schema refusal from current host permission.

    Node.js 24 · New npm folder · Three pinned packages
04 / TECHNICAL TRACK

Evaluation, deployment and operations

Agentic engineering

Inspect traces, bound the worker, recover failed work and transfer a system another engineer can operate.

WHAT YOU MAKEA release and recovery plan backed by local lab evidence
Start with 1017 connected lessons
102 · Explore the guided lesson sequence
  1. 01Evaluations: how do you know it works?45 min
  2. 02Observability: what happened inside the request?40 min
  3. 03Evaluate the task, the trajectory and the failure100 min
  4. 04HTTP and FastAPI: your first request40 min
  5. 05Networking: from a service name to a response40 min
  6. 06Deploy an agentic workflow with an operator exit160 min
  7. 07Earn multi-agent complexity and hand over the system70 min

Academy lessons open with a provisioned training seat.

More labs in this track 7

Choose the boundary you want to practise. Each walkthrough includes setup and local commands.

  • Evaluate the task, the trajectory and the failure

    Regression comparison lab

    Compare the working reference with broken candidates and retain missing-case evidence.

    Python standard library · Keep the workflow file beside it

  • Earn multi-agent complexity and hand over the system

    Shared delegation coordinator

    Trace retries, shared attempt limits and late results after manual closure.

    Node.js 24 · Supplied event replay

  • Deploy an agentic workflow with an operator exit

    Receiver commit and replay

    Compare a duplicate-producing receiver with an atomic effect and receipt, then retry a lost response.

    Python standard library · Owned child-process exits

  • Evaluate the task, the trajectory and the failure

    Agent trace and payload boundary

    Keep bounded observability events, then inspect optional real OpenTelemetry spans without a collector.

    Core: Python 3.11+ · Standard library · Optional SDK extension

  • Deploy an agentic workflow with an operator exit

    Agent canary evidence

    Compare candidate/control segments, retain missing evidence and stop an aggregate-green regression.

    Python 3.11+ · Standard library · Supplied observation packets

  • Deploy an agentic workflow with an operator exit

    Private inference capacity

    Budget weights, KV cache, context and output reservations without claiming measured GPU capacity.

    Python 3.11+ · Standard library · No GPU or model

  • HTTP and FastAPI: your first request

    Your first HTTP request with FastAPI

    Follow real requests through routing, validation, Python functions and four distinct responses.

    Python 3.11+ · Isolated environment · Supplied pinned requirements

BRING IT TOGETHER

The client problem is the through line.

Use the tools to deliver an agreed outcome, explain the tradeoffs and hand over work the receiving team can maintain.

Sign in to the learning workspace

The DEPLOY Method

Carry the work from ambiguity to ownership.

DEPLOY is the field rhythm. Each move produces evidence before the engineer receives more authority.

  1. 01

    Diagnose

    How does the work actually move, especially where the documented process breaks?

    Mission charter, stakeholder map, and observation boundary

  2. 02

    Establish

    What is the smallest safe wedge that can produce decision-grade evidence of workflow impact?

    Outcome contract and wedge decision record

  3. 03

    Prove

    Can the smallest wedge meet the mission contract under representative, edge, adversarial, and changing conditions?

    Runnable wedge and provider-neutral behavioral contract

  4. 04

    Land

    Can the proven wedge live inside enterprise identity, data, control, and operating boundaries without losing its evidence?

    Production-shaped enterprise integration slice

  5. 05

    Operate

    Can users adopt the system while the team detects, contains, and learns from behavior in production?

    Production-readiness and staged-rollout record

  6. 06

    Yield

    Can the named owner operate, change, and recover the system—and what permitted learning should compound?

    Owner-accepted transfer package and owner-executed change/recovery record
Read the complete field standard

The formation sequence

Eight missions. Built to be completed in motion.

12-week cohort formation of structured formation can be followed, where authorized, by supervised residency. The runway is policy—not a time-to-competence claim.

  1. 97 min

    Diagnose the work, establish the outcome

    Run evidence-led discovery, reconstruct the real workflow, and convert ambiguity into a falsifiable outcome contract without prescribing a model first.

  2. 87 min

    Scope and architect the mission

    Translate the outcome contract into a thin, reversible system design with explicit deterministic, model, tool, human, and operating boundaries.

  3. 103 min

    Land inside enterprise systems

    Build a production-shaped vertical slice through identity, data, legacy interfaces, tools, human gates, telemetry, and recovery.

  4. 109 min

    Engineer provider-portable AI systems

    Own the behavioral contract, isolate provider capabilities behind adapters, and make model selection an evidence-backed engineering decision.

  5. 114 min

    Operate an evaluation and reliability control system

    Convert the outcome contract into layered evals, detect a behavior regression, and make release decisions from failure evidence rather than demos.

  6. 102 min

    Bound authority and earn authorization

    Threat-model the landed system, design least-privilege controls and evidence, and defend an authorization recommendation under attack.

  7. 119 min

    Operate adoption, rollout, and incidents

    Move a proven system through staged authority, measure accepted outcomes and adoption, and lead incident response under changing evidence.

  8. 103 min

    Yield ownership and compound field learning

    Transfer operation and change to a named client owner, defend the complete mission, and turn permitted field evidence into reusable product learning.

Open the complete mission sequence

Inside every module

The same five moves, every time.

Read, check, practice, evidence, close — every authored module below is built the same way. The figures beneath are a direct count over the curriculum, not a claim about it.

  1. 01

    Read

    The deep dive opens with a mechanism diagram, then works through the lesson section by section.

  2. 02

    Check

    Knowledge checks follow the reading. Every option explains itself, right or wrong, from the lesson text.

  3. 03

    Practice

    One interactive exercise — classify, order, spot, or decide — rehearses the method, with every choice explained.

  4. 04

    Evidence

    The module's artifact is submitted and reviewed against a rubric published before the work begins.

  5. 05

    Close

    A closing summary names what the reader carries forward and the handoff to the next module.

TRACKS
8
MODULES
49
MECHANISM DIAGRAMS
49
KNOWLEDGE CHECKS
147
PRACTICE EXERCISES
49
HOURS OF READING
6
HOURS OF ARTIFACT WORK
55

Inside the work

Instruction ends where judgment begins.

Inspect one real lesson structure and one production-shaped exercise. Provisioned workspaces remain private.

Sample lesson

Prompting as specification

Prompts are programs in natural language. Write them with intent, context, constraints, and acceptance tests.

Inspect lesson architecture

Field exercise

Hold or ship the support agent

A support team wants to move a ticket agent from model M1 to M2 before Monday. You have one release packet, a segmented eval, and pressure to take the cheaper model live.

Try a bounded assessment

The evidence bar

Completion is visible. Entrustment is earned elsewhere.

The current authored standard contains ten conjunctive evidence gates. Training progress cannot substitute for production experience or independent human judgment.

10

evidence gates

  1. 01Unassisted baseline work sample
  2. 02Discovery and outcome contract
  3. 03Enterprise integration slice
  4. 04Provider-portable system
  5. 05Evaluation and reliability control
  6. 06Security and authority review
  7. 07Production, adoption, and incident command
  8. 08Supervised field mission
  9. 09Ownership transfer and field learning
  10. 10Integrated qualification defense

The standard is inspectable. Qualification issuance opens once a new qualification instrument is preregistered; the current instrument is not preregistered.

Inspect the qualification constitution

LockedIn Labs FDE

Build something another team can trust.

Request consideration Discuss an enterprise cohort