Articles and guides for forward deployed engineers

Focus your reading

Read client/server roles, requests, responses, and statelessness.

Put it to work

Trace a failed integration across each hop; distinguish transport, authentication, and application failures.

Before you start

Skip if you can already explain headers, methods, status codes, cookies, and proxy behavior.

Content checked. Reviewed 2026-10-31

Focus your reading

Use EXPLAIN basics, EXPLAIN ANALYZE, and caveats; compare estimates with actual rows.

Put it to work

Annotate one slow plan with the first bad estimate, expensive operator, and a measured change.

Before you start

Requires SQL joins and indexes. ANALYZE executes the query; use a safe test environment.

Content checked. Reviewed 2026-10-31

Focus your reading

Read the customer-workaround example and discussion of generalization.

Put it to work

Write the underlying job, smallest acceptable solution, and conditions that would justify productizing it.

Before you start

Practitioner account; one successful workaround is not a rule against building features.

Indexed content checked. Reviewed 2026-10-30

Focus your reading

Focus on integrating data, working alongside users, and the tension between local usefulness and generality.

Put it to work

Describe one belief about the workflow that changed after observing it.

Before you start

A personal retrospective, not a universal operating model or independently audited case.

Content checked. Reviewed 2026-10-30

Focus your reading

Read the ambiguous-chunk example, hybrid lexical/semantic retrieval, and reranking evaluation.

Put it to work

Compare recall on exact identifiers versus semantic queries, holding the evaluation set fixed.

Before you start

Requires a retrieval baseline. Published gains apply to the tested setup, not every corpus.

Content checked. Reviewed 2026-10-31

Focus your reading

Read just-in-time retrieval and techniques for long-running tasks, including compaction and structured notes.

Put it to work

Record which evidence a task needs, when it is loaded, and what must survive compaction.

Before you start

Prefer a reproducible failure before changing memory architecture; test information loss after compaction.

Content checked. Reviewed 2026-10-31

Focus your reading

Compare routing, chaining, orchestrator/worker, evaluator/optimizer, and autonomous agent patterns.

Put it to work

Explain why a deterministic implementation or fixed workflow is insufficient before adding an agent loop.

Before you start

Patterns remain useful; the 2024 article explicitly notes that its tooling has evolved.

Content checked. Reviewed 2026-10-30

Focus your reading

Read tool selection, namespaces, response content, pagination, and held-out evaluation.

Put it to work

Compare two tool contracts on task success, invalid arguments, redundant calls, latency, and token cost.

Before you start

Requires traces of real tasks. Do not score success only by matching one prescribed sequence of tool calls.

Content checked. Reviewed 2026-10-31

Focus your reading

Start with error analysis, sampling traces, stale datasets, and validating an LLM judge against expert labels.

Put it to work

Create a labeled failure taxonomy and report judge false accepts and false rejects on held-out examples.

Before you start

Avoid turning suggested sample counts into universal rules; coverage and failure cost determine the needed evidence.

Content and selected detailed answers checked. Reviewed 2026-10-31

Focus your reading

Read tasks, trials, graders, transcripts, outcomes, and repeated-run metrics.

Put it to work

Build a release report with outcome checks, multiple trials where appropriate, and unresolved failure slices.

Before you start

Keep grader validation separate from agent evaluation; aggregate success can hide critical failures.

Content checked. Reviewed 2026-10-30

Focus your reading

Focus on least privilege, deny-by-default, per-request checks, and authorization regression tests.

Put it to work

Create an access matrix with allowed and denied actions across users, tenants, resources, and exports.

Before you start

Authentication success is not authorization. Test object-level checks on the server.

Content checked. Reviewed 2026-10-31

Focus your reading

Paper section 2: relation tuples, the “new enemy” problem, and consistency tokens.

Put it to work

Specify revocation and content-update ordering for a retrieval service; demonstrate a forbidden read stays forbidden.

Before you start

Study the invariant without assuming you need Google-scale infrastructure. See case C17 for the transfer question.

Paper and authorization-consistency section checked. Reviewed 2026-10-31

Focus your reading

Read confused-deputy attacks, per-client consent, token audience checks, and the prohibition on token passthrough.

Put it to work

Draw who issues each token, its audience, which component validates it, and the downstream authorization boundary.

Before you start

This link is draft guidance; implement against your chosen released protocol version and identity provider.

Content checked; draft is not a pinned stable specification. Reviewed 2026-10-31

Focus your reading

Read indirect injection, least privilege, action checks, and the stated limits of model-based guardrails.

Put it to work

Threat-model a retrieved document that asks for a write or data export; enforce permissions outside the model.

Before you start

Filters and a second model do not establish a complete security boundary. Test the actual action path.

Content and defense limitations checked. Reviewed 2026-10-31

Focus your reading

Read timeout selection, retry amplification, backoff, and jitter.

Put it to work

Give a dependency a time budget and bounded retry policy; identify the one layer responsible for retrying.

Before you start

Retrying a failed-looking write can repeat its effect; pair with the idempotency resource.

Content checked. Reviewed 2026-10-30

Focus your reading

Read request identifiers, server behavior, and retryable operations.

Put it to work

Specify key scope, request-parameter matching, concurrent duplicate handling, and result retention.

Before you start

A deduplication key is part of a contract, not a blanket guarantee about downstream systems.

Content checked. Reviewed 2026-10-30

Focus your reading

Inspect load patterns, input/output length distributions, time-to-first-token, and end-to-end latency.

Put it to work

Report latency distributions against request rate and task length; state model, hardware, and workload.

Before you start

README reviewed, tool not run. Synthetic benchmark throughput is not application quality.

GitHub README checked; software not executed. Reviewed 2026-10-31

Focus your reading

Read organizational adoption, experiment interactions, and skepticism toward surprising results.

Put it to work

Specify the comparison, unit of assignment, guardrails, and what would invalidate a claimed lift.

Before you start

Large-scale web experiments do not directly fit every enterprise pilot; state limits when randomization is infeasible.

Author page and paper checked. Reviewed 2026-10-31

Focus your reading

Look at how training is organized around designing, building, and deploying agents.

Put it to work

Create a role-specific onboarding plan and test whether an operator can diagnose and escalate a failure unaided.

Before you start

A company account of training design, not evidence that a particular course causes better outcomes.

Indexed content checked. Reviewed 2026-10-30

Focus your reading

Focus on customer-defined evaluations and replacing repeated custom implementations with reusable capability.

Put it to work

Identify one repeated field constraint and propose a shared change with a second-deployment test.

Before you start

Company-reported examples. Do not infer universal ROI or staffing ratios from them.

Content checked. Reviewed 2026-10-30

Focus your reading

Review lifecycle commands, evaluation, infrastructure, CI/CD, observation, and upgrade paths.

Put it to work

Create a vendor-neutral package checklist and have a second engineer deploy from the documented inputs.

Before you start

Google-specific reference, not a required stack. It succeeds Agent Starter Pack, now in maintenance mode.

GitHub README checked; software not executed. Reviewed 2026-10-31

Focus your reading

Day 37: the runbook test, security review, observed stalls, and findings routed into fixes.

Put it to work

Have a peer follow the runbook without coaching. Record blocked steps and give each finding an owner.

Before you start

The published reference run used agents as simulated participants without a real customer cloud account. It does not establish human usability or production deployment; the lesson itself acknowledges this.

Lesson and reference-run limitations checked; course and deployment not independently executed. Reviewed 2026-10-31