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
All guides
GuideLevel · Deep#Agents#LangGraph#Reliability

Your agent can resume. Can it avoid doing the same thing twice?

Durable execution, checkpoints and external side effects explained through an approval-and-resume workflow.

DCDariusz Czechowicz23 Sept 20263 min readPractical guide
21
Guide
Source-checked Sources & corrections
4 practical sections · checklist · sources

Practical guide. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.

For: Developers building agents that run beyond one request

Before you start: Basic state machines and transaction semantics

⚡ TL;DR — the 30-second version

  • Recovery and exactly-once effects are different guarantees.
  • Checkpoint mode is a workload-specific choice.
  • Revalidate authority when an agent resumes.

Long-running agents need to survive provider errors, process crashes and human delays. Persistence helps, but it creates an important distinction: restoring internal state is not the same as knowing whether an external side effect already occurred.

Imagine an agent that creates a support ticket, then crashes before saving the ticket ID. On resume, a naive workflow creates the ticket again. The model did not cause this bug; the execution contract did.

“A checkpoint remembers where you were. An idempotency contract decides what may happen again.”
— Editorial interpretation

Field guide / 01

01 / Separate intent, execution and confirmation

Represent a proposed action, its approval, its execution attempt and its confirmed result as different states. Keep the action's stable identifier outside model-generated text. Store enough information to reconcile an ambiguous outcome with the downstream service.

For an operation that supports idempotency keys, reuse the same key across retries of that logical action. For one that does not, design an application-level deduplication or reconciliation process rather than promising exactly-once delivery.

Field guide / 02

02 / Choose checkpoint guarantees deliberately

LangGraph documents exit, async and sync durability modes. Exit persistence is cheaper but loses intermediate recovery on a crash. Async writes can overlap execution but leave a small loss window. Sync waits for each checkpoint before proceeding.

These are trade-offs, not a universal setting. A read-only research job can tolerate different loss behavior than a workflow changing customer records.

Checkpoint modeTrade-offQuestion to ask
exitLess intermediate persistenceCan the whole run safely repeat?
asyncOverlap execution and writesWhat happens if the last checkpoint is lost?
syncWait for persistenceIs the extra latency worth the recovery guarantee?

Field guide / 03

03 / Resume with the same authority

Persisted state is not a permission grant. Revalidate the user and action scope when resuming. A role may have changed while the agent waited; a document may have been deleted or a customer's consent revoked.

Bind approval to an immutable action payload or its reviewed version. If a model revises the action after resuming, it needs another authorization decision.

Field guide / 04

04 / Test at the dangerous boundaries

Terminate the process just before an external call, just after the call, and before the confirmation is stored. Test expired credentials and a revoked approval. These cases reveal more than a long happy-path demonstration.

Report attempted actions, confirmed actions and reconciled uncertain actions separately. If the system cannot prove what happened, surface that uncertainty to an operator instead of asking the model to invent a reassuring summary.

Before you ship

0/5

Use this as a review aid. Checks reset when you leave this page.

Put it into practiceInspect an agent's permission boundary

Sources & verification

Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.

  1. 01LangGraph — orchestration, persistence and human reviewLangChain↗
  2. 02LangGraph — checkpointers and durability modesLangChain↗
  3. 03MCP 2026-07-28 — authorizationModel Context Protocol↗
Editorial policy and correction log →
#Agents#LangGraph#Reliability#State