Doso — Concept Overview

Doso combines a workflow graph for execution with a knowledge graph for durable context. The two graphs solve different problems and meet through typed contracts, provenance, and review.

What you build remains yours:

  • Knowledge and workflow graphs persist without vendor lock-in.
  • Connections use your Composio account, your own integrations, or a paired desktop runtime that runs host CLIs.
  • The runtime can host locally, on your server, or in any cloud and country.
  • Inference uses any provider and any model you configure.

Two kinds of graph (keep them distinct)#

Workflow graph (orchestration)Knowledge graph (memory)
NodesAgent jobs with contractsEntities (people, orgs, places, events, artifacts, …)
EdgesData dependencies / control flowTyped predicates (S–P–O triples) with provenance
PurposeFan out work, verify, mergePersist facts beyond a context window; multi-hop answers

They compose: the workflow graph is how agents run; the knowledge graph is the durable substrate they read/write and fact-check against.

Glossary#

  • Node (workflow): one bounded job — one agent, explicit input, schema-validated output.
  • Edge (workflow): a real dependency; data must cross. “And then” without data is a fake edge.
  • Contract: input shape + output schema + single responsibility; what makes a node wireable.
  • Diamond: fan out → reduce (code) → synthesize (judgment) — the default serious topology.
  • Verifier / checker: edge-gated node that tries to kill a finding before it reaches the answer.
  • Anchor: non-negotiable ground truth (tests passed, real outcomes) — topology without anchors is self-deception.
  • Entity / triple: knowledge-graph node and typed relation; descriptions aid resolution.
  • Resolution: clustering surface forms into a canonical entity (description-aware, not string-similarity alone).
  • Provenance: which document/chunk produced a node or edge — required for cited answers.

How the concepts fit together#

  1. The knowledge pipeline extracts, resolves, and assembles durable facts with provenance.
  2. Agentic memory gives workers and verifiers a shared substrate that survives individual sessions.
  3. The workflow graph coordinates bounded jobs, parallel branches, joins, and approval gates.
  4. Ontology and grooming turn the underlying graph into useful product views without creating a second source of truth.
  5. The architecture keeps these concerns separate while exposing them through one product.