SeekLook within intelligence, not only at its outputs.
OUR MISSION
Seek within.Evolve beyond.
SeedSky seeks within intelligent systems to understand how intelligence works.
We turn that understanding into a foundation for clearer control, continuous learning, and lasting evolution.
Infrastructure for understanding, controlling, and evolving intelligence from within.
SEEDSKYINTELLIGENCEhow intelligence unfolds
Context
Process
Behavior
Outcome
01 · SEEKObserve
02 · SEEDUnderstand
03 · EVOLVEControl
04 · BEYONDEvolve
SeedTurn what we learn into a durable data foundation.
EvolveTranslate understanding into measurable change.
SkyExpand what intelligent systems can become.
THE SEEDSKY THESIS
Self-improvement begins with self-observation.
Intelligence is not only defined by inputs and outputs, but by the process in between.
- 01Interpret
- 02Transform
- 03Respond
FROM OUTSIDEOutput tokens
FROM WITHINNeuron-level mechanism
Both describe the same task. One is orders of magnitude larger.
↺
What the system changes becomes what we observe next — at mechanism level.
This inner process is where intelligence becomes observable, understandable, controllable, and improvable.
THE SEEDSKY LOOPFrom inner understanding
From inner understanding
to compounding improvement.
SeedSky connects what happens within an intelligent system to what it changes in the world.
SEEK WITHIN
EVOLVE BEYOND
Observe
Make the inner process and its consequences visible.
Understand
Reveal patterns that explain capability, limits, and change.
Control
Turn understanding into precise, measurable intervention.
Evolve
Return evidence to the system so each cycle can improve the next.
WHY MECHANISM CHANGES THE SEARCH
Existing self-improvement searches behaviour: change something, run the task, keep whatever scored better. Reading the mechanism turns that blind search into a directed one.
Targeted
Internal evidence points at the capability in question, so the search starts from a prior instead of a guess.
Efficient
Mechanism narrows the space before the first trial, so fewer runs are needed to reach the same frontier.
High-throughput
With the direction known, candidates can be explored in parallel instead of one attempt at a time.
RETURN SIGNALEvery outcome becomes evidence for the next cycle.
OBSERVE → UNDERSTAND → CONTROL → EVOLVE ↺AVAILABLE NOWControl how hard
Control how hard
the model thinks.
Coding agents ship with a handful of thinking levels and no way to choose between them per task. cct adds the levels that were missing and puts the choice in your hands.
WHAT WE FOUND
Mechanistic analysis of the model's internals, driven through a self-evolving search pipeline, surfaced the channels that carry reasoning effort. We extended the three official thinking levels to five, and found two settings that sit off the official ladder entirely.
icct is a launcher that wraps the coding agent you already use. It starts the agent and rewrites the requests it sends, so every task runs at a thinking level you chose. It does not retrain anything, it does not send your work anywhere, and every change it makes can be inspected or switched off.
The run command ships with the first release.
REASONING CONTROLA clear objective for every task
LOCAL RUNTIMECURRENT TASKCross-file refactor with tests
model.run
MIND POWER
Flex spans these five efforts — medium and xhigh are ours
Better task outcomes for fewer tokens than the level above it.
MID
MID
STD
◆
APPLIED PROFILEvalue · 2026.08Bounded transform · unchanged-request fallback
THE EVIDENCEWorking points beyond
Working points beyond
the official ladder.
Mechanism makes the search targeted, efficient, and high-throughput, so it reaches regions blind search never will. value and deeper live there — the chart shows what they return against the official low / high / max.
Cost per successfully completed taskLocked holdout · replace with verified results
SeedSky profileProvider tier
The ladder, completed
The official ladder jumps low → high → max. In the gaps between, the search found medium and xhigh — two real, stable steps.
value — off the ladder
A working point no official level reaches: better task outcomes for fewer tokens.
deeper — past max
Where the official max stops thinking, deeper keeps going.
Search with a map
Mechanistic evidence points straight at the channels that carry thinking effort. The search is targeted — it starts from a map, not a guess.
Fewer runs to the frontier
The map cuts most of the space before the first trial. That efficiency is why the frontier is reached in far fewer runs.
Sweep in parallel
With direction fixed, candidates run side by side. High throughput means every sweep covers more of the thinking space — and more key regions get found.
RESEARCH FOUNDATIONBuilt from research
Built from research
into how intelligence works.
The runtime grows from a broader research program on inner process, measurable behavior, inference-time control, and the relationship between mechanism and outcome. The first product is one proving ground, not the whole horizon.
Inner processHow intelligence unfolds between context and outcome
BehaviorWhat changes across tasks and conditions
EvidenceWhich patterns predict capability and failure
ControlWhich interventions reliably improve results
THE MISSIONSeek within. Evolve beyond.
THE MISSION
Seek within.Evolve beyond.
We are building the data and control infrastructure through which intelligence can understand its own operation and evolve beyond today's frontier.