Can Your Free Automation Survive a Retry?

In n8n or Dify, route a synthetic intake batch to a test queue and replay a lost acknowledgement. Measure executions and duplicate writes before choosing a builder or enabling the live trigger.

Make the write succeed before the test reports failure

Start with an inbound inquiry classifier that proposes a department and writes to a test queue. Use invented inquiries and derive a stable event ID from each input record so retries carry the same identity. In the test harness, let the write complete and then simulate a lost acknowledgement. Replay the same event through the builder and inspect the queue. The acceptance test is one queued item for that ID and an observable record of both attempts. Use a sink that enforces uniqueness on the ID, or an atomic upsert that preserves the first accepted classification. For example, if the first attempt routes an inquiry to billing and a second model call chooses sales, a duplicate item or a changed department exposes a replay failure even when the builder reports a successful second run.

Timeouts, retries, and backoff with jitter | AWS Builder Center

Read the execution counter during the replay

For the n8n hosted option, the supplied pricing excerpt shows a Starter trial and a paid Starter plan measured in workflow executions. Use the entitlement shown for your account when deciding whether your test fits. Run a synthetic batch of ten new inquiries, then resubmit two of those IDs through the trigger. Record the before-and-after usage counter, model calls, and unique queue items. If the counter rises by 12 while the queue gains ten items, budget 1.2 executions per unique inquiry for that sample. Measure an internal retry separately because resubmitting through a trigger and retrying within a run can show up differently in your records. If you test Dify Cloud for the same job, repeat the fixture using its account allowance and compare the observed usage per completed inquiry in each builder’s billing unit.

n8n Plans and Pricing - n8n.io

Start with the simplest pattern that fits

Before comparing builders, decide what kind of system you actually need. One documented framing distinguishes workflows, where LLMs and tools are orchestrated through predefined code paths, from agents, where the LLM dynamically directs its own process and tool usage. The same guidance recommends finding the simplest solution possible and only increasing complexity when needed, noting that for many applications a single optimized LLM call with retrieval and in-context examples is enough.

That matters because agentic systems often trade latency and cost for better task performance. A lead qualifier that asks the same five questions in order is a workflow. A support assistant that decides which tool to call next is an agent. Label your task before you shop, because the label determines which kind of builder you need.

Illustrative check: if you can draw the task as a numbered list of steps with no branching except "if answered, continue," treat it as a workflow candidate rather than an autonomous agent.

Building effective AI agents \ Anthropic

Build a prototype and define pass-fail signals

A practical first build is one repeated task with a binary outcome. Define a pass condition, run the prototype against a fixed set of inputs, and record failures before expanding scope. Documented platform capabilities support this: Dify combines AI workflow, RAG pipeline, and agent capabilities with model management, and its RAG pipeline covers document ingestion through retrieval with text extraction from PDFs and other common formats. That lets you ground answers in your own material rather than general model knowledge.

Observable success signals: identical inputs produce factually consistent answers, retrieved content visibly matches the source document, and the agent stays within its defined goal. Failure signals: answers drift from uploaded material, irrelevant documents are retrieved, or the run stops partway without a clear error. Keep the test inputs fixed so regressions are visible after any prompt or model change.

Illustrative example only: upload three policy PDFs on day one and treat retrieval accuracy, not conversational polish, as the acceptance metric.

GitHub - langgenius/dify: Build Agentic workflows, RAG pipelines, with rich AI model and tool support on one collaborative workspace. Deploy on cloud, VPC, or self-hosted, so teams move from prototype to production without rebuilding the stack. · GitHub

Understand what an execution counts as

Usage limits are where free tiers and paid plans diverge, so learn the unit of billing before you scale. On n8n Cloud, an execution is a single run of an entire workflow, regardless of step count or data volume, which the pricing page contrasts with per-operation or per-task billing in other products. The Starter plan is listed at 20 euro per month billed annually with 2.5K workflow executions with unlimited steps, and Pro at 50 euro per month billed annually. These are the plan rates exactly as published.

Map your own volume before choosing. If your agent handles 50 conversations daily, that is roughly 1,500 runs monthly (illustrative arithmetic, not a vendor measurement). Compare that against the published execution allowance and the plan's other limits, since Starter also lists fixed constraints on shared projects and concurrency.

Paid plans may include, as n8n's page states, AI models without API keys, so verify whether model access is included or billed separately.

n8n Plans and Pricing - n8n.io

Schedule recurring runs and check timezones

Workflows that should fire on a cadence need a trigger. The n8n Schedule Trigger node runs workflows at fixed intervals and times, similar to Cron, and supports second, minute, hour, day, week, month, and custom Cron intervals. Two operational notes are documented: a workflow using the Schedule node must be saved and published, and the node relies on a timezone setting. n8n uses the workflow timezone if set, otherwise the instance timezone. The self-hosted default is America/New_York; n8n Cloud tries to detect the owner's timezone at signup and falls back to GMT.

A recurring follow-up email or a nightly lead report fits this pattern. Success signal: runs appear at the expected local time after a publish and timezone check. Failure signal: a silent no-show because the workflow was saved but not published, or runs landing at unexpected hours because the instance timezone differs from yours.

Illustrative troubleshooting: set the workflow timezone explicitly before scheduling across regions.

Schedule Trigger | Nodes | n8n Docs

Sources