05 / Landscape
Own the semantics. Reuse the machinery.
Yukh is being tested against the agent ecosystem with no protected components and no preferred external winner. A useful result may be a pivot, a replacement, multiple supported adapters, or less Yukh code.
Research status / 24 August 2026. This page is a dated qualification view, not an adoption decision, and it does not supersede accepted RFCs. RFC-0003 remains the current reference architecture.
What changed
The architecture review is no longer only documentary. goose and Hermes both passed the first executable runtime-substrate gate. The stricter Gate A2 now requires a two-sided Yukh/host policy-composition proof through supported public seams; it is defined but candidate adapters have not yet executed it. No reference host is selected.
Memory has also been separated from orchestration. TencentDB Agent Memory is now the primary Track C2 shared/team-memory qualification candidate, while OpenHuman remains an orchestration/workflow challenger and a memory benchmark rather than a combined default platform.
The architectural test
Yukh should own only governance and interoperability semantics that must survive a change of model, host, runtime, workspace or vendor. Agent hosts, sandboxes, workflow engines, memory systems and human interfaces should remain replaceable machinery whenever external implementations preserve those semantics.
Current qualification posture
| Component / project | Posture | Why it matters | Disqualifier |
|---|---|---|---|
| Yukh MCP | Keep semantics; qualify composition | Cross-runtime capability policy, decision and evidence boundary | Becoming a second approval engine beside every host |
| Yukh Coordination | Keep kernel; integrate surfaces | Claims, conflict, handoff, attribution and replay | An external protocol proves equivalent semantics with less custom code |
| Yukh Projects | Narrow to adapter | GitHub Projects implementation of accepted work state | Treating GitHub Projects itself as a universal Yukh layer |
| goose | Host candidate — substrate PASS / A2 not executed | ACP + MCP host with broad composability | Yukh composition requires internal patching, global auto mode or broader authority |
| Hermes Agent | Host candidate — substrate PASS / A2 not executed | Provider-neutral persistent host with ACP, MCP, approvals and multiple execution backends | Host memory or internal APIs become required Yukh semantics |
| Apache Maka | Durable-execution candidate | Event-sourced AgentRun and execution/evaluation facts | Required records leak private reasoning or authority cannot remain external |
| OpenHands | Execution-isolation candidate | External sandbox and remote execution machinery | Isolation requires adopting unrelated platform authority |
| OpenHuman | Orchestration challenger | Checkpointed workflows, approvals and agent fleets; memory remains a benchmark concern | Broad platform coupling or workflow state becomes Yukh authority |
| TencentDB Agent Memory | Shared-memory candidate — Track C2 | Team memory assets across agents: Chat Memory, Skill, Wiki and CodeGraph | Memory status/ACL/proxy semantics leak into Yukh evidence or project authority |
| Buzz | Workspace candidate | Human-agent rooms, identity/attribution and collaborative interaction | Its event substrate changes Yukh authority or handoff semantics |
| Codex / Claude Code | Participants, not dependencies | Agent runtimes that should be able to participate through public seams | Yukh core contracts become vendor/runtime-specific |
What Yukh may actually need to own
| Semantic concern | Stable Yukh contract? | Implementation direction |
|---|---|---|
| Participant identity / attribution | Yes | Yukh-visible participant contract; runtime/workspace identities remain adapters |
| Claim / conflict / handoff | Yes | Yukh Coordination today; external transports and workspaces may carry it |
| Capability policy / decision | Yes | Yukh MCP semantics composed with host-native permissions |
| Evidence correlation | Yes | Operational facts must survive host/runtime replacement without private reasoning |
| Agent host / loop | No | goose, Hermes, Codex, Claude Code and other replaceable hosts |
| Sandbox execution | No | OpenHands and host/runtime-native sandboxes |
| Durable AgentRun facts | Candidate stable contract | Apache Maka first qualification target |
| Durable orchestration / workflow | Probably not core | OpenHuman and other workflow engines |
| Persistent agent memory | No, except boundary contracts | TencentDB Agent Memory; Hermes/OpenHuman as benchmarks |
| Accepted work state | Yes, abstractly | GitHub Projects through yukh-projects today; other adapters later |
| Human collaborative workspace | No | Buzz, Matrix, IDEs, web clients |
Most plausible pivots
- Projects becomes explicitly an adapter. The durable semantic is accepted work state; GitHub Projects is the current implementation, not necessarily a universal pillar.
- Yukh never builds an agent host. goose and Hermes have both earned further qualification; neither has yet passed Yukh-specific host composition.
- Memory becomes external team machinery. TencentDB Agent Memory tests whether reusable Chat Memory, Skills, Wiki and CodeGraph can remain contextual without becoming authority or evidence truth.
- Orchestration remains external. OpenHuman tests whether checkpointed workflows and approvals can be consumed without turning workflow state into Yukh authority.
- Buzz becomes a collaborative surface. Yukh Coordination can remain the semantic kernel while Buzz, Matrix or another workspace carries human-agent interaction.
- OpenHands supplies isolation. Yukh policy can authorize a bounded operation while external execution machinery performs it and returns correlated evidence.
- A Yukh repository can disappear. Prior investment is not a reason to preserve a component that an external project makes redundant.
Qualification tracks
- Track A / Agent host: goose vs Hermes. Runtime-substrate PASS for both; Yukh host-composition A2 next.
- Durable execution: qualify Apache Maka for AgentRun / execution facts and evidence.
- Orchestration: qualify OpenHuman workflow/checkpoint/approval boundaries independently from memory.
- Track C2 / Shared memory: qualify TencentDB Agent Memory against memory ≠ authority, accepted state or evidence truth.
- Workspace: map Buzz interaction and identities onto Yukh Coordination without semantic loss.
- Execution isolation: execute one bounded Yukh capability through OpenHands and return correlated evidence.
- Accepted state: re-evaluate yukh-projects as the GitHub adapter for an abstract accepted-work-state contract.
The dated research record and governing issues carry the evidence. External projects are never considered selected merely because they appear on this page.
Architecture review / #56 Host qualification / #58 Shared memory qualification / #63