Skip to main content

    How tool chaining compares

    Several platforms offer folding, docking and ADMET models. What differs is how a job runs and who connects one job's output to the next: an agent, a pipeline you build, or code you run.

    conceptsorchestratorcomparison

    Tool catalogues overlap. Tamarind Bio lists 300+ molecular design tools and Phylo's Biomni Lab lists 300+ databases, software systems and analytical tools, each by its own count. Several of them fold, dock or score ADMET with the same published models .bloom* runs. So "which tools are there" rarely decides the choice. What decides it is how a job runs, and who carries one job's output into the next.

    How a job runs on .bloom*

    Ask Ivy a question and it plans the run, submits each job, and feeds each job's output into the next as a job reference. You don't download a structure to upload it again for docking. The chain is built from your question, so changing a step means naming a tool or a parameter. See how the orchestrator picks tools.

    Any tool also runs on its own: one job from the terminal, a curl request or a studio project, with no agent involved. Each step is a row in the run's record. signed-in runs from the web app, the terminal, curl or MCP leave a tamper-evident record you can check, and so does a guest chat, from the browser tab it is open in, while the guest session lasts.

    How other platforms connect the steps

    Each row below is built from that vendor's own public pages, as read in October 2026. "Not advertised" means their materials don't list it, not that it is absent. The full side-by-side is on /platform/comparison.

    PatternWho connects the stepsExamples
    One job per submit, pipelines you buildYou: each tool is a separate submit (submit, poll, download), and chaining is a pipeline template you build and re-runTamarind Bio
    An agent writes and runs codeThe vendor's agent, writing code in a workspace: its cloud, or compute you connectPhylo / Biomni Lab (its cloud workspace, plus nf-core pipelines for sequencing); Claude Science (your machines, cluster or Modal account)
    The vendor's agent, inside its appIts agents, within the app's own scopeBenchling AI (steps beside your ELN); Amazon Bio Discovery (antibody models, then partner labs); BenchSci EMET (skills over its knowledge graph)
    Workflows you build in its softwareYou, in a desktop app or PythonSchrödinger (Maestro or its Python API); OpenEye / Cadence (Orion floes)
    Your own agent calls the toolsYour agent, one tool call at a time, over MCP, a CLI or PythonToolUniverse

    What the difference changes

    • Moving outputs between steps. Where you build the pipeline, you map each tool's output fields to the next tool's inputs, once per template. Where an agent writes code, the glue is code you can read but also have to check. On .bloom*, Ivy passes job references between steps.
    • Changing the plan mid-run. A pipeline template runs as built; a different question means editing it. A question-driven chain changes with the question.
    • What is kept. Tamarind doesn't advertise a record of the run. Phylo keeps the code, sources and execution details behind each analysis, and Claude Science keeps the code, environment and conversation behind each result. On .bloom* each step is a row in a hash-chained record.

    When a pipeline you build fits better: the same fixed steps, re-run over large batches, where you want every parameter pinned in a template you own. That is what a pipeline builder is for.

    Was this page helpful?

    On this page

    On this page

    Loading...