Harness before prompt
Tools, context, permissions, memory, and stop conditions form the system. The model is one component inside it.
* Independent AI research company
We build from
the loop out.
AgentOrch.ai explores difficult AI systems problems by making agents the primary builders. More than 90% of each research build is executed inside agent harnesses, from discovery and implementation to evals and iteration.
AGENT SHARE90%+
HUMAN INTENTHIGH
MODERESEARCH
LIVE ORCHESTRATION MAPHARNESS_01 / NOMINAL
[ 01 / THESIS ]
Agent-native, not agent-assisted.
The agent is not a feature.
It is the production system.
We design the harness before we optimize the prompt. We build loops before we add handoffs. We instrument evals before we trust output. The result is a research practice where agents do the bulk of the work and humans hold the thesis, boundaries, and taste.
Loops before handoffs
Useful autonomy comes from observable cycles of action, critique, repair, and evidence, not a longer one-shot instruction.
Evals before opinions
Every meaningful claim needs a trace, a metric, or a reproducible test. Vibes are a signal, never the benchmark.
[ 02 / PROJECTS ]
Research you can run.
Tools, not
just claims.
Our projects turn research questions into usable systems. Every release is a working artifact, an evaluation surface, and an input to the next loop.
LIVE PROJECT / 001OPEN SOURCE
RAG Cost
Calculator
Model the cost and operating profile of retrieval-augmented generation systems. Compare architecture choices before infrastructure decisions become expensive.
- Model comparison
- Cost scenarios
- Architecture planning
Open the project +
NEXT PROJECT SLOT
The research queue is open.
New experiments and utilities will appear here as they graduate from the harness.
STATUS / FORMING[ 03 / RESEARCH SURFACE ]
Problems worth
building agents for.
Our research surface is organized around the infrastructure required for capable, economical, and verifiable agent systems.
Autonomous research harnesses
Agent systems that can frame a question, assemble context, run experiments, and preserve a legible chain of evidence.
- orchestration
- context
- tooling
Self-improving knowledge systems
Retrieval and reasoning loops that measure their own misses, repair weak context, and compound what they learn.
- rag
- memory
- evals
Agentic cost intelligence
Instrumentation for the real economics of autonomous work: quality, latency, token spend, retries, and human escalation.
- telemetry
- economics
- routing
Multi-agent verification
Critic, judge, and challenger patterns that make long-running agent loops more observable, defensible, and safe.
- verification
- judges
- control
[ 04 / OPERATING LOOP ]
Research that
compounds.
Each run leaves the next run stronger. Findings become fixtures. Failures become evals. Useful traces become reusable context. The harness gets better while the research moves forward.
10%Human
Thesis, constraints, judgment, taste.
90%+Agents
Research, code, tests, analysis, documentation.
AGENTORCH / EXPERIMENT LOOPCONTINUOUS
- 01FrameNEXT
- 02ScaffoldNEXT
- 03RunNEXT
- 04EvaluateNEXT
- 05CompoundREPEAT
> evidence persisted> eval delta +3.8%> next run queued
[ 05 / KNOWLEDGE OUTPUT ]
Build in public.
Write with evidence.
Projects produce traces, ideas, and methods worth sharing. The journal captures the build; papers preserve the research.
JOURNALOPENING SOON
Blogs and build notes
Field notes from inside the harness: architecture decisions, failed loops, cost discoveries, evaluation patterns, and opinions earned through building.
FIRST SERIESWhat changes when agents build 90% of the system?
Follow the research log +PAPERSIN PREPARATION
Research papers
Methods, benchmarks, and findings made reproducible. Papers will connect the thesis to evaluation design, experimental evidence, and reusable artifacts.
PLANNED NOTEMeasuring agent share across a production research loop
Review the research surface +[ 06 / FOUNDERS ]
Built by researchers who build.
Human conviction.
Agent-scale execution.
FOUNDER / 01NPProduct strategy / agent economics / systems architecture
Narsimha Rao Polisetty
Co-founder
FOUNDER / 02HJResearch direction / evaluation / intelligent systems
Dr. Hemant Joshi
Co-founder
AGENTORCH.AI / RESEARCH NODE