FlowDrop AI BenchA living benchmark of AI content workflows

Architectures

What each of the ten benchmark cells does, drawn with one shared vocabulary, and how the cells relate: which pairs differ by exactly one decision.

Every cell runs the same prompt on the same pages. What differs is the architecture around the model: what it is given (a URL, the HTML, or Markdown), whether it runs once or in a loop, which tools it holds, and who owns them. The labels used everywhere on this site read input the model sees → who does the work → tools.

The ladder

Controls, no model

The floor and the reference. Free, deterministic, and the yardstick for time and structure.

One model call

A fixed graph with a single model call in it. The graph decides the order of work; the model does one job.

Read them in pairs

Each pair differs by one decision, so the gap between the two rows in the matrix or the cost figures is the price or the value of that decision.

B2 vs B3Who converts. B2 makes the model transcribe HTML; B3 converts in a node and spends the model on redaction only.
B3 vs B4Who sequences. Same two jobs; in B3 the graph orders them, in B4 the agent decides whether and when to call the converter.
B5 vs B6Which runtime. Same autonomy, same tools, same prompt; B5 is a FlowDrop ReAct agent, B6 a Drupal AI Agent.
B5 vs B7Tool shape. B7 adds one fused fetch-and-convert tool, so the page text need not pass through the model.
B7 vs B8Tool ownership. Same tools; in B7 the agent holds them, in B8 they are nodes of the parent workflow handed in as an argument.
B8 vs B9A critic. B9 is B8 with a review loop: a second model call checks the draft and can send it back, up to three times.

Every cell

inputoutputDeterministic nodefixed behaviourModel callthe model decidesAgentmodel chooses what runsToola tool the model may calltool callHTMLdata, named by its payloadthe agent loop, until doneparent workflowthe graph that owns the tools
The vocabulary every figure below uses.

B0 Orchestration floor

The URL goes straight to the output node. No fetch, no conversion, no model.

URLURL, untouchedOutputnothing runsthe URL
What it tests. How much of every run is FlowDrop overhead: validation, pipeline creation, one dispatch, persistence. Every other cell's time is read against this floor.

On the other pages this cell is B0 URL → output, nothing runs. Its results across models.

B1 Deterministic reference

Fetch the page, convert the HTML to Markdown with a node. Nothing is redacted.

URLURLFetchHTTP nodeHTMLTo Markdownhtml_to_markdown nodeMarkdownMarkdown, unredacted
What it tests. The structure the model cells are scored against, byte-identical every time and free. Because it redacts nothing, it shows what the redaction step adds.

On the other pages this cell is B1 URL → fetch → to-Markdown, no model. Its results across models.

B2 Raw HTML to one model call

Fetch the page, then hand the raw HTML to one model call that converts and redacts at once.

URLURLFetchHTTP noderaw HTMLModel callconvert and redactMarkdown
What it tests. The least build effort of any cell. Its input tokens scale with the size of the page markup, not the size of the article. Read against B3.

On the other pages this cell is B2 HTML → LLM. Its results across models.

B3 Markdown to one model call

Fetch, convert to Markdown with a node, then spend one model call on the redaction only.

URLURLFetchHTTP nodeHTMLTo MarkdownnodeMarkdownModel callredact onlyMarkdown
What it tests. The hybrid: the graph does the mechanical half, the tokens buy judgement instead of transcription. Compare with B2 for the price of conversion in the model, with B4 for who decides the order. Read against B2, B4.

On the other pages this cell is B3 Markdown → LLM. Its results across models.

B4 Drupal AI Agent given the HTML

Fetch in the graph, then pass the HTML to a Drupal AI Agent that owns an HTML-to-Markdown tool and decides for itself whether to call it.

URLURLFetchHTTP nodeHTMLDrupal AI Agentgiven the HTMLModel callredact, may convertmay skip the toolhtml_to_markdownreturns MarkdownMarkdown
What it tests. Same two jobs as B3, but the sequencing is the model's call rather than the graph's. The gap between B3 and B4 is the price of letting the model decide. Read against B3.

On the other pages this cell is B4 HTML → AI Agent + to-Markdown tool. Its results across models.

B5 FlowDrop ReAct agent given the URL

Only a URL goes in. A FlowDrop ReAct agent holds a fetch tool and a to-Markdown tool and reasons its way to the answer; the page text returns to the model after the fetch and leaves it again for the converter.

URLURL onlyFlowDrop ReAct agenttools in its own toolboxModel callplan, call tools, redactreason · act · observe · repeatFetchreturns HTMLTo MarkdownHTML in, Markdown backMarkdown
What it tests. The fully autonomous end of the ladder with nothing about the order of work in the graph. B6 is the same shape on Drupal AI Agents; B7 and B8 change one thing each. Read against B6, B7, B8.

On the other pages this cell is B5 URL → ReAct agent + fetch, to-Markdown tools. Its results across models.

B6 Drupal AI Agent given the URL

Only a URL goes in. A Drupal AI Agent retrieves the page with its own http_fetch tool, converts with html_to_markdown and redacts. No FlowDrop HTTP node is involved.

URLURL onlyDrupal AI Agenttools in its own configurationModel callplan, call tools, redactreason · act · observe · repeathttp_fetchreturns HTMLhtml_to_markdownHTML in, Markdown backMarkdown
What it tests. The Drupal AI Agents peer of B5: same autonomy, a different agent runtime, so the pair separates the architecture from the implementation. Read against B5.

On the other pages this cell is B6 URL → AI Agent + fetch, to-Markdown tools. Its results across models.

B7 ReAct agent with a fused fetch-and-convert tool

B5 plus one more tool, URL-to-Markdown, which fetches and converts in a single call. The agent can take the URL to Markdown without the page text ever passing through the model.

URLURL onlyFlowDrop ReAct agentone more tool than B5Model callplan, call tools, redactreason · act · observe · repeatFetchreturns HTMLTo MarkdownMarkdown backURL to Markdown+URL in, Markdown backMarkdown
What it tests. What a better-shaped tool is worth: same agent, same prompt, one tool that removes a round trip of the whole page through the context window. Read against B5, B8.

On the other pages this cell is B7 URL → ReAct agent + fetch-and-convert tool. Its results across models.

B8 ReAct engine, tools owned by the parent workflow

The same three tools as B7, but they are nodes of the parent workflow, handed to a bare ReAct engine as an execution argument. The engine's own toolbox is empty.

parent workflowowns the toolsReAct engineown toolbox emptyModel callplan, call tools, redactreason · act · observeFetchreturns HTMLTo MarkdownMarkdown backURL to MarkdownURL in, Markdown backtools, passed in as an execution argumentURLURL onlyMarkdown
What it tests. Who owns the tools: the agent or the graph around it. Same reasoning loop as B7; the change is where tools live, run and are observed. Read against B7, B9.

On the other pages this cell is B8 URL → ReAct engine, tools in parent graph. Its results across models.

B9 Reflexion engine with a critic, tools in the parent

B8 with the Reflexion engine: a second model call reviews the draft Markdown against the fidelity and redaction rules and sends it back for revision, up to three times, before it is emitted.

parent workflowowns the toolsReflexion engineown toolbox emptyActorplan, tools, draftCriticreview the draftdraftREVISE, up to 3 timesFetchreturns HTMLTo MarkdownMarkdown backURL to MarkdownURL in, Markdown backtools, passed in as an execution argumentURLURL onlyMarkdown
What it tests. What a critic buys. Same tools and prompt as B8; the only addition is the review loop, so its cost and its effect on silent failures are read directly against B8. Read against B8.

On the other pages this cell is B9 URL → Reflexion + critic, tools in parent graph. Its results across models.