Target Architecture¶
Coactra's target public surface is Team · Agent · Workflow - small composition primitives for policy-aware AI workloads. The library wraps mature runtimes (LangGraph, Temporal, Graphiti, mem0, LiteLLM, MCP, A2A) rather than reimplementing them. Coactra owns the policy, tenancy, scope, capability roster, and connector contracts that sit above those engines.
Core decisions:
- One public assembly door:
from coactra import Team, Scope, Policy, Workflow Team(scope=..., policy=...)owns agent registration, skill catalogs, workflow catalogs, and routingteam.add_agent(...)is the only endorsed application construction path for runtime agents- External packages can expose plain install functions such as
install_hermes(team); do not add a public plugin protocol until real integrations need one Workflow(steps=[...])routes by broadrequires_skilldomains plus optionalrequired_tags; directagent=pinning is an override- Workflow routing and execution both go through Team policy checks
- Approval resumes carry structured
ProofBundleevidence, not plain text claims coactra.agent.recipes.code_change(...)is an optional helper for implement/verify/review, not a replacement for dynamic workflow inductionairemains an internal engine layer, not a public package identitymemoryandworkspacestay adapter-backed capabilities governed by Team policyDepartment/Companystyle hierarchies are optional composition layers, not mandatory core concepts- Auth follows OAuth 2.1 / MCP gateway / A2A Agent Cards standards
The authoritative source for the full architecture - the composition model, build order, brownfield notes, and supporting specs - is the system vision document:
design/2026-06-06-coactra-vision.md
The external-review refinements that tighten standards alignment, Team-first assembly, gateway-primary auth, structured Skills, and workspace security: