Skill-collaboration layer: a user's own agents collaborate to improve their skills (in-taOS)

[jaylfc]

Vision (Jay 2026-06-14)

A cross-agent learning system where agents learn from each other. A dedicated skills A2A channel/bus where agents collaborate: share lessons learned, memories, and skill-file improvements, so good knowledge propagates across the fleet instead of being trapped in one agent.

Why it fits taOS now

• The A2A bus already exists and agents already coordinate on it (@taos / @taOSmd / @Hermes). A skills (or learning) channel is a natural addition.
• Goal-driven capability planning + local-first knowledge model (read-only guides + agent-authored supplements) #898 already proposes agent-authored supplementary guides (personal + shareable, layered over the read-only canonical guides). This is the distribution + collaboration layer for those: an agent that learns "FLUX Q3 fits a 12GB 3060, Q4 OOMs" publishes that supplement; others subscribe and benefit.
• The agent-audit/observability layer (Universal control plane: live queue + full-trace observability/audit that every interaction flows through #896) is where you'd review shared lessons before they spread (governance, no bad lessons propagating).

Shape

• Skills A2A channel: agents post lesson/memory/skill-improvement events (structured: capability, hardware tier, claim, evidence, provenance/author).
• Subscribe + merge: an agent pulls relevant shared supplements (layered over canonical guides, never overwriting them, per Goal-driven capability planning + local-first knowledge model (read-only guides + agent-authored supplements) #898), scoped by capability/hardware so it only adopts what's relevant.
• Curation/governance: review queue (human or trusted-agent) before a shared lesson is promoted from "one agent's note" to "fleet supplement" -- ties to Universal control plane: live queue + full-trace observability/audit that every interaction flows through #896 audit + the "why did the agent do that" governance need.
• Reputation/provenance: each lesson carries who learned it + evidence, so conflicting lessons can be ranked, and a bad one rolled back.
• Shareable like apps/themes/packages (publish to community), per the taOS distribution model.

Tie-ins

• Goal-driven capability planning + local-first knowledge model (read-only guides + agent-authored supplements) #898 (agent-authored personal/shareable guides -- the artifact this shares).
• Universal control plane: live queue + full-trace observability/audit that every interaction flows through #896 (control plane / audit -- review + governance of shared lessons).
• The A2A bus + the canonical-skills read-only model (Canonical taOS-operations skills: one source of truth, framework-portable injection, permission-bound #595 concept).
• Memory system (taOSmd) -- shared memories overlap; design WITH @taOSmd.

Notes

Big, novel, and on-brand for the agent-audit/learning differentiator. Brainstorm WITH @taOSmd (memory side). Vision capture; do not build now.

[jaylfc]

Scope clarification (Jay 2026-06-14)

This is a product layer inside a user's taOS, for the user's own deployed agents to collaborate on improving their skills -- NOT the dev/build agents (@taOS/@taOSmd). Reframe the issue with that scope:

• A user runs several agents in their taOS (different frameworks/roles). When one learns something useful -- a better way to do a task, a skill-file improvement, a memory/lesson -- it can share it with the user's other agents, so the user's fleet collectively gets better at the user's work over time.
• Mechanism (in-instance): a shared skill-collaboration channel/store scoped to the user; agents publish learnings (provenance + evidence) and subscribe/merge what's relevant to their role/capability, layered over the read-only canonical guides (Goal-driven capability planning + local-first knowledge model (read-only guides + agent-authored supplements) #898), never overwriting them.
• User governance: the user sees and approves what propagates between their agents (review queue), surfaced via the control plane (Universal control plane: live queue + full-trace observability/audit that every interaction flows through #896). The user owns the shared knowledge; it stays in their perimeter.
• Optional community sharing: a user can publish a proven skill improvement to the store like apps/themes/packages (per Goal-driven capability planning + local-first knowledge model (read-only guides + agent-authored supplements) #898's shareable supplements), but the DEFAULT is private to the user's instance.
• Builds on: the skills system (skill_exec/skills.py), per-agent memory, Goal-driven capability planning + local-first knowledge model (read-only guides + agent-authored supplements) #898 (agent-authored guides = the artifact shared), Universal control plane: live queue + full-trace observability/audit that every interaction flows through #896 (review/governance). The A2A bus is the transport but the LAYER is the product feature: user's agents teaching each other.

So: a user's agents form a little learning team that compounds. Design WITH @taOSmd (memory side).