Overview
Apache Maka’s architecture treats durable facts as the authority for long-running agent work, then derives bounded views for interfaces, model context, recovery, graph scheduling, evaluation, and privacy-sensitive operations. It follows one request through the Runtime Event Log, one-writer Runtime Host, T1/T2 side-effect boundaries, validated context projections, and evidence-bound continuation. Its value is a clear way to distinguish ownership, committed outcomes, projections, and unknown states; its limits include no universal exactly-once effects, bit-exact provider replay, workspace snapshots, or universal sandbox coverage, and the project remains Apache Incubating without an Apache release.
Learning outcomes
- Durable RuntimeEvents—not transcripts, callbacks, or screens—anchor what an agent can later know.
- A one-writer Runtime Host gives each State Root clear execution and recovery ownership.
- T1 marks possible dispatch and T2 makes a result durable; neither creates exactly-once external effects.
- Compaction, Graph, Eval, and privacy policies remain bounded projections or controls, not replacements for Runtime authority.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What does Apache Maka treat as the source of truth for agent work?
Apache Maka treats the Runtime Event Log as the canonical record of semantic interaction facts; interfaces, model context, recovery, and other consumers derive bounded views from it.
Why does Apache Maka use one Runtime Host for a State Root?
Apache Maka uses one writer Host per State Root so clients share execution and recovery authority without creating competing histories; different Sessions can still run concurrently.
Can Apache Maka guarantee exactly-once tool effects or exact replay?
No. T1/T2 expose the dispatch-to-result gap, but arbitrary shell commands, remote APIs, and child processes are not exactly-once; semantic replay is stronger than provider-native replay, and neither promises bit-exact wire replay.
How does Apache Maka recover when a tool result is missing?
It repairs the old Run, classifies the operation from durable boundaries, reconciles external evidence where available, and creates a fresh continuation only after safety checks pass; otherwise it parks the work as unresolved.
Does Apache Maka’s local workspace mean no data leaves the machine?
No. Maka stores runtime data locally and ships no cloud sync, but a chosen cloud provider or integration may receive data; local storage is not a universal transmission or sandbox guarantee.