Independent. Engineering-led. We're building the intelligence layer between AI coding agents and the code they work on. Talon, the context layer, is live today. We use it ourselves on every working day.
We're building the intelligence layer between AI coding agents and the code they work on. Talon, the context layer, is live now: it serves the right context to your agent before each run, so the same model ships in fewer turns for measurably less spend.
Context intelligence and steering for your coding agent. Talon predicts what each task will need, scores what actually matters in your codebase, and steers the agent clear of the loops that burn turns and tokens — a prediction and heuristics engine, not just a graph. Same model, fewer turns, less spend, measured under a sealed benchmark.
How it works →Two more layers of the stack, both in active development. We'll open them up when they're ready to measure.
Session analysis — see where a run went sideways: context bloat, wasted turns, scope drift, turn by turn.
The coherence engine for long-running agents. Continuity · Attention · Economics. Your agent at turn 400 is the agent that learned everything at turn 5 — at a tenth of the spend.
Cogmeta started from a small observation: paying Anthropic by the request and watching agents grep their way through code that already has a perfectly good structural index sitting right there, unused. The metadata exists. It just isn't being used. Cogmeta is the layer that uses it.
The journey ahead is to leverage our discoveries in agent control, optimisation and context, augmented with a multi-dimensional memory layer, to realise the full benefit loop optimisation can deliver.
20+ years across startups, global banking, telco, and government — shipping and operating distributed systems at scale.
AI coding agents write code well. They find the right code to write less well — and they forget what they learned by next session. That gap costs real money — and we believe it's solved between the model and the codebase, not inside the model.
Rediscovery loops and missed dependencies — they get worse as the codebase grows, and they only show up on your bill.
Bigger windows and smarter prompts plateau fast. The fix lives at the layer deciding what the model sees — not inside it.
A layer between the model and your code that decides what it sees. Same model, same prompts — less cost, fewer wrong turns, measured against a paired baseline in /research. We publish the losses next to the wins.
Please reach out if you have questions, wish to discuss anything further.