swarm
Multi-agent research with a live audit trail: watch the agents work, and trace every claim back to a verified source.
Multi-agent systems fail enterprise review because nobody can see inside them. What did the agents do? What did they spend? Which claims did they invent? A governance team is right to reject a black box that produces a confident report. swarm starts from the opposite premise: the audit trail is the product.
A root agent turns the prompt into a tree of sub-tasks, then stops. No tokens are spent until a person confirms the plan — the spend gate is structural, not a setting.
Sub-agents fan out under a concurrency cap and stream findings live over server-sent events into a force-directed graph — the hierarchy, every finding, and cross-links whenever two findings cite the same source.
A verifier re-fetches every cited source to confirm the claim it supports. The writer has no web access at all; it builds only from findings that survived verification.
A reviewer checks the draft against its provenance and flags fabrications at three severity levels. On a pass, the run collapses into an auditable artifact: the interactive graph, a written report, and the raw structured data.
Stack: TypeScript and Node throughout — Express backend, server-sent events for streaming, Cytoscape for the live graph, Claude for every agent role, strict lint, and a full test suite.
Most agent frameworks report cost after the run. swarm inverts that: the orchestrator plans, then stops, and whoever owns the budget sees how many agents will run and how deep the tree goes before the first token is spent. The gate is not a setting you can forget to enable; it is the only path through.
Rather than asking one model to be accurate, the pipeline separates powers. Researchers find; a verifier re-fetches 100% of cited sources; the writer has no web access, so it cannot invent a URL; a reviewer audits the draft against provenance at 3 severity levels. A fabrication has to slip past all four, and each fails differently.
Every dispatch, finding, and verification streams to the browser the moment it happens, so the graph your team watches is the system's complete history. When the run ends, that graph collapses into the artifact your auditors keep.
Every cited source is re-fetched before a claim ships, and a reviewer flags fabrications at three severity levels. A completed run yields the reasoning graph, a written report, and the raw JSON. Built and run by us, at our own expense.
The orchestration layer runs inside your cloud, wrapped around your own tools and data sources. Your team gets the graph view for audit, and the confirm gates are wired to whoever owns the budget.