Pentad Labs Research Notes

Research notes from Pentad Labs on the architecture of WunderOS and related topics, including harness and context engineering, agentic memory, deterministic runtimes, supervisory control, and the systems theory underneath an operating system for agents.

Each note states a claim and the reasoning for it. They are dated and numbered, reflective of their place in an ongoing stream of engineering and inquiry.

  1. PLRN-000 The law of agentic friction 31 May 2026
  2. PLRN-001 Coordination and memory are the same problem 5 May 2026
  3. PLRN-002 One ledger, two runtimes: determinism across Zig and the BEAM 5 May 2026
  4. PLRN-003 Sometimes the smart move is the shorter string 5 May 2026
  5. PLRN-004 Control agents without modifying them 5 May 2026
  6. PLRN-005 Every layer reduces to one evaluator 5 May 2026
  7. PLRN-006 Provenance is a byproduct of evaluation, not instrumentation 2 Jun 2026
  8. PLRN-007 Agent memory: never delete, decay and retract 2 Jun 2026
  9. PLRN-008 Agent tool-call integrity without an LLM 2 Jun 2026
  10. PLRN-009 Recording the parse, re-deriving the plan 6 Jun 2026
  11. PLRN-010 A standing query is an agent's native shape 12 Jun 2026
  12. PLRN-011 When agents fail mid-plan, a saga tree is the path back 16 Jun 2026
  13. PLRN-012 Every subtree runs as the system: one vocabulary, one boundary, at any scale 18 Jun 2026
  14. PLRN-013 No view from nowhere for enterprise memory 19 Jun 2026
  15. PLRN-014 Authority without assertion for agent execution 22 Jun 2026
  16. PLRN-015 Extract first, canonicalize later, for agent memory 22 Jun 2026
  17. PLRN-016 No memory of its own: governing a visiting agent on sovereign data 30 Jun 2026
  18. PLRN-017 Memory for agents is missing who said what 6 Jul 2026
  19. PLRN-018 Two clocks, one honest agent-OS record 8 Jul 2026
  20. PLRN-019 A Datalog of, by, and for agents 10 Jul 2026
  21. PLRN-019A Gleamalog by example: the mortgage casebook 10 Jul 2026
  22. PLRN-020 PlatypusDB: the database agents actually want 20 Jul 2026
  23. PLRN-021 An autonomic agent OS is closed control loops, not automation 27 Jul 2026
  24. PLRN-022 Use the cheapest model that preserves the answer 28 Jul 2026
  25. PLRN-023 Memory compaction for agents has an information-theoretic floor 4 Aug 2026
  26. PLRN-024 Put agents on every desk and firm output may not move 12 Aug 2026
  27. PLRN-025 No NLP in an agent OS hot path 22 Aug 2026
  28. PLRN-026 Sleeping agent fleets are cheap. Waking them must be fast. 12 Sep 2026
  29. PLRN-027 A harness can't keep an agent's future-dated promises. An agent OS can. 20 Sep 2026