Talk to one model.
Questions, careful analysis, math, and code share the same conversation.
Open Echo ↗TRACER / RESEARCH LAB
LATEST
Echo
Intelligence is
a system property.
We study how coordination creates capabilities no model has alone, then turn that work into systems that deliver better performance per token. Echo is our latest product.
YBacked by Y CombinatorEcho is now live: one model name and one OpenAI-compatible endpoint for chat, code, and agents.
ECHO / ONE ADAPTIVE SYSTEM
LISTENING
REQUEST ARRIVING
01 / 04
ECHO / ADAPTIVE FREQUENCYAWAITING REQUEST
INTELLIGENCE ALLOCATED0%
QUICKDEEP
Tracer is a YC-backed research lab. Echo turns our work on coordinated intelligence into a product people can use.
Our thesisThe thesis
Because... 1 + 1 > 2.
An orchestra is not a louder violin. It can do what no instrument can do alone.
Violin
Cello
Horn
Drums
Coordinate
Symphony
Emergent whole
A capability created by coordination
Echo
One endpoint.
Many kinds of work.
Echo is powered by adaptively coordinated open-weight models behind one stable interface. It adapts the capabilities and computation to each request.
from openai import OpenAI
echo = OpenAI(
base_url="https://echo.tracerml.ai/v1",
api_key=ECHO_API_KEY
)
Keep your OpenAI client.
Set Echo's base URL, model name, and API key. Echo adapts behind the same stable interface.
Read the API docs ↗Use Echo across the agent loop.
Point OpenCode and other compatible harnesses at Echo. The harness keeps its tools and loop; Echo adapts inside each model call.
Connect OpenCode ↗For teams running AI in production
Better performance.
Lower inference cost.
Echo is our latest product. We also work directly with teams to improve performance per token through intelligence allocation and coordination.
Book a 30-minute inference review For teams evaluating token efficiency, model performance, or inference unit economics.Spend intelligence where it matters.
Apply the right amount of computation to each kind of work, preserving depth for the tasks that earn it.
Make the system stronger than one model.
Coordinate complementary capabilities around the outcome instead of forcing every request through the same shape.
Prove the economics end to end.
Evaluate quality, latency, token use, and cost against your actual workload and constraints.
Research lineage
Built in public.
Echo grows from our work on learned allocation, interpretability, and auditable deployed systems.
TRACER OSS
A trace-driven, parity-gated inference system and the beginning of our research program.
GitHub Opens in a new tabTRACER paper
Our first work on trace-based cost efficiency for LLM classification: how lightweight models can learn recurring decisions while preserving a measured quality gate.
Read the paper ↗From efficiency to emergence
Our current investigation: coordinated specialists, adaptive intelligence, and capabilities of the whole.
Research in progressResearch programs
We study systems,
not isolated scores.
Our work sits where machine learning, complex systems, and production infrastructure meet.
Coordinated intelligence
When can specialists create a capability that no individual model has?
Adaptive inference
How should a system allocate computation as the work changes?
Interpretable orchestration
Can allocation decisions be inspectable, stable, local, auditable, and contestable?
Agentic systems
Which parts of a trajectory require reasoning, and which can become learned reflexes?
Working manifesto
TRACER
MMXXVI
The next leap may not be
a larger mind.
It may be a better-organized one.
We believe intelligence can emerge from coordination.
We study when adaptive coordination can become self-organizing.
That efficiency and capability can reinforce one another.
That systems should reveal how their decisions are made.
And that discovery often begins where the whole becomes stranger and more capable than its parts.
Questions, answered plainly
What Echo is.
What it is not.
What does the adaptive display represent?
The visual represents how Echo adapts its execution strategy as tasks change. It is not a live execution trace, literal architecture diagram, or menu of user-selectable modes.
Can Echo run inside OpenCode?
Yes. Configure Echo's base URL, model name, and API key in OpenCode. The same OpenAI-compatible endpoint also works in other compatible harnesses. Read the integration guide.
How is Echo evaluated?
We publish frozen comparisons with matched denominators, stored rows, hashes, losses, and explicit limitations. The evidence is inspectable in the read-only Echo Eval Observatory.
What does Tracer mean by coordinated intelligence?
We study whether specialized capabilities, learned allocation, verification, and adaptive computation can create a whole that is more capable or efficient than its strongest isolated component. Until it does, emergence remains a hypothesis to test.
Work with Tracer
More intelligence.
Less wasted compute.
Bring us the workload, constraints, and inference bill. We will help you find where better allocation and coordination can improve both performance and cost.