Compounding intelligence · the QSKL knowledge stack
Intelligence that compounds.
A chat request becomes an objective, and the agent works the loop. What it learns does not have to evaporate when the run ends: every run can leave behind a skill, and every skill can earn its way to a crystal.
Included in QuoxCOREFree, self-hosted, nothing to buy.
How it compounds
Five steps, the same loop every time.
Nothing here is a one-off pipeline. It is a loop a run goes through every time, and the loop is what makes the intelligence compound: what one run leaves behind, the next run can start from.
A request becomes a run
You ask in chat. QuoxCORE turns the request into an objective and hands it to an agent as a governed, agentic run: the same routed, policy-checked hand-off every agent on the platform goes through.
The engine nudges skill creation
As the agent iterates, the engine periodically proposes turning the working procedure into a skill: versioned and scoped to your org, starting life in an experimental edition. 49 skills ship seeded across 12 categories today.
Every invocation is telemetered
Each call records a session id, tokens used and a drift score. A weekly judge scores clarity, completeness, recency and drift; a daily cron cools skills nobody is calling, ageing them from active to cooling to archive candidate.
Skills improve through patches
Three consecutive successes propose a patch capturing the winning tool sequence. It has to clear the skill's own tests and a crystal regression suite, sits inside a 60-minute auto-revert window and a rate limit, and passes a static-analysis firewall and a safety judge before experimental can become stable.
Proven runs become crystals
A run that keeps working gets frozen for deterministic replay and signed with the org's crystal signing key before it can go canonical. Next time the same task appears, the crystal replays instead of the agent re-deriving it: the same outcome, fewer tokens, evidence attached.
In the running product
The loop lives in AI Studio.
This is a real capture of AI Studio, not a mockup. The Crystals Library sits beside Agents, Workflows and Skills: the same tiers the loop above walks through, one screen away from where you already work.
- 1One workshop, twelve tilesAssistants, agents, workflows, prompts, skills and crystals sit in a single studio rather than spread across separate settings pages.
- 2Agents, where a run startsSpecialist agents configure their own tools, capabilities and autonomy: the run stage the skill nudge fires from.
- 3Crystals Library, marked newA proven run freezes here as a signed, replayable crystal, a separate tier from the skill list beside it.
- 4Skills, marked newThe versioned, org-scoped procedures the engine proposes every fifteen iterations live in this tab.
- 5Skill ReviewsWhere the weekly four-dimension judge and the daily decay cron leave their read on each skill.
- 6QuoxChat, marked newThe chat surface that starts the loop: a request typed here becomes the objective an agent turns into a run.
Captured from the real AI Studio grid, not a mockup.
The crystal tier
Two kinds of crystal.
A crystal forms one of two ways: caught in the moment, or written up afterwards. Both end at the same place, a canonical, signed, replayable run.
Task-time candidate
Drafted while a run is still live: the engine watches the loop, drafts a candidate from what it observed, and proposes it with a diff of what was learned. No action is needed for the draft to appear, but nothing is promoted without an explicit approval.
Post-hoc canonical
Written up after the fact: pick a run that worked, name it, describe it, and promote it deliberately. The traditional path for turning something you already know works into something the library can retrieve.
Every crystal moves through the same four states, draft, candidate, canonical, deprecated, whichever path it started on. Read the full lifecycle on the crystals concept page →
Governed, not assumed
An evidence layer, not a black box.
A skill that improves itself is only trustworthy if every step it takes is on the record. That is the design, not an afterthought bolted on once something learned the wrong lesson.
01 · evidencedEvery step is witnessed
The nudge, the patch proposal, the firewall verdict, the safety judge’s read, the promotion itself: each one emits a VOLT envelope. Nothing in this loop happens off the record.
▸ VOLT envelopes · witnessed02 · org-scopedYour skills stay yours
Every skill row carries your org id. A request for a skill in another org’s library returns 404, not 403, so cross-org existence never leaks either.
▸ quox_skills · org-scoped03 · gatedPromotion has to earn it
The static-analysis firewall and the safety judge sit inside the promotion path, not next to it. A skill cannot reach stable without clearing both, and a pending safety review blocks the move outright.
▸ firewall + safety judge · gatedWhere it stands
What's proven, what's not yet.
- 49 seed skills live across 12 categories, versioned and scoped to your org
- Real invocation telemetry: session id, tokens used, a drift score per call
- A weekly four-dimension judge and a daily decay cron, both default-on
- A patch flow gated by the skill’s own tests, a crystal regression suite, a 60-minute auto-revert window, a 10-patches-per-agent-per-24h rate limit, a firewall and a safety judge
- A separate crystal tier for deterministic replay, distinct from the skill tier
- A cross-org marketplace for skills: the publications table and the publish path exist, there is no subscribe flow, curator index or Ed25519 signing on that path yet
- Vector or semantic skill search: filtering is SQL WHERE only today
- Multi-modal skill composition: the schema has a column for it, nothing in the seed catalogue uses it
- Per-action telemetry: invocations emit typed rows, but the action_log column was never added to the table
None of this has run at production scale yet: the platform reports zero live skill invocations at the time of writing. The telemetry, judge and decay infrastructure are real and running, they simply have not been exercised under real load.
Go deeper
The protocol and the platform.
Ready when you are
Start a run, see what it leaves behind.
QuoxCORE ships with the skill and crystal stack built in: no separate install, no extra key.