Agent platform

Status: complete (published 2026-08-19). Scope finalized at Batch D kickoff with maintainer POV; see brief.md, findings.md, vendors.md, sources.md.

Scope inventory (v0)

• Build: SDKs and frameworks, planner and multi-agent topologies, routing tiers (skill selector, classifier, LLM router, fallback), prompt and context management, compaction
• Run: session and event-log architecture (brain/hands/session decoupling), tiered memory services, guardrail integration points, per-run budgets and circuit breakers, human-in-the-loop primitives (approval queues)
• Govern: agent catalog with owner and risk tier, skill and tool catalogs, model catalog, publication workflow (build, approve, stage, rollback, retire), versioning, developer portal
• Sourcing: build on primitives vs managed platforms (Frontier, Agent 365, Agentforce) vs hybrid; portability and exit strategy

Challenged-default candidates

Multi-agent frameworks vs a single well-built loop; managed coworker platforms vs composable primitives. Each track proposes its final list at kickoff.

Files

brief.md, findings.md, vendors.md, sources.md are created from ../_TEMPLATE/ at kickoff.

---

Source: research/R07-agent-platform/README.md in the evidence repository behind this site.