DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///DROPGoogle launches Gemini 3.8 Flash TTS and Flash-Lite TTS///DROPGemini Omni brings 1080p generation to Google Vids///BREAKClaude Opus 5.5 is here: 1M context, $4 input / $20 output///BREAKOpenAI adds GPT-6 Sol and Luna to the GPT-6 family///PAPEROpus 5.5's system card documents gains—and evaluation limits///BREAKGrok 4.7 launches at $2 input / $6 output per million tokens///DROPGrok Voice Transcribe 2.0 targets difficult real-world audio///TRENDGoogle steers new Gemini API projects to 3.8 Flash or 3.5 Flash-Lite///
DARC.DEV

[ 03 — Advanced · 6 lessons · ~4.7h with exercises ]

Agents & Automation

Build an inspectable agent and then give it durable state, scoped tools, human approval and enforceable spending limits. Learn where deterministic workflows are the better choice.

← All learning paths

Your progress

0/6

Mark completion after doing the exercise. The checklist records your self-assessment, not a certificate. Stored for this browser only.

Module 01

Start with one loop

A clear task and a small toolset beat unexplained autonomy.

~45 min · read + exercise

Build a bounded agent

Define success and failure exit states.

Exercise

Sketch a read-only research agent with finite retries and a budget.

Done when: Every state can end, pause or retry within a documented limit.

Read the lesson

~45 min · read + exercise

Instructions, approvals and sandboxes

Assign a job to each enforcement layer.

Exercise

Test three denied operations in a disposable repository.

Done when: The tool layer blocks the action independently of model cooperation.

Read the lesson

Module 02

Connect safely

Protocols and visual tools still need authorization.

~55 min · read + exercise

MCP authorization and tool trust

Distinguish token audience, user authority and tool annotations.

Exercise

Draw the trust map for one local and one remote MCP integration.

Done when: You can show where downstream credentials live and why token passthrough is disallowed.

Read the lesson

~40 min · read + exercise

Draft, validate, approve, execute

Build a workflow around a narrow model step.

Exercise

Design a support-response automation that never sends without exact-content approval.

Done when: Replaying the same input cannot send the same message twice.

Read the lesson

Module 03

Operate long-running work

Persistence and cost controls must survive failure.

~55 min · read + exercise

Checkpoint and resume

Separate recovery from exactly-once side effects.

Exercise

Walk through crashes immediately before and after an external write.

Done when: You have a reconciliation path for an unknown delivery outcome.

Read the lesson

~40 min · read + exercise

Enforce gateway budgets

Test actual rejection instead of trusting a configuration file.

Exercise

Plan a staging test with a deliberately small budget and a restart.

Done when: The cap is still enforced after restart and fallback obeys the same data policy.

Read the lesson