PROJECTSAKTHILong-Horizon Intelligence
We build intelligent systems that can plan, adapt, and learn across extended time horizons, turning complex, multi-step goals into reliable outcomes.
Intelligence is a long game: planning, feedback, improvement.
AGI Research Insights
Latest synthesis of public forecasts and community evidence on AGI timelines/capabilities.
AGI probability by 2040
Central tendency across major forecasts.
Earliest credible prediction year
Optimistic—but sourced—first-demo estimates.
Pre-2060 predictions
Share of researchers expecting AGI before 2060.
Predictions analyzed
Combined sample used for our internal outlook.
We track signals from scaling trends, benchmark inflection points, and evaluation studies.
Our Mission
Advance the science of long-horizon autonomous systems, capable tool-use, robust alignment, and verifiable safety, while delivering efficient, domain-specific models that solve real tasks today.
Research Pillars
Long-Horizon Planning
Algorithms that plan and execute multi-step tasks over time.
Tool-Use & Program Synthesis
Systems that choose, compose, and even generate tools.
Alignment & Robustness
Keep systems safe and value-consistent under distribution shift.
Evaluation Science
Hard, reproducible measurement for capability and safety.
Efficient & Domain-Specific Models
Smaller SLMs and specialists that matter in production.
Methods & Infrastructure
Our stack is built for reproducibility and scale.
Public Evals & Data Releases
We maintain rotating, leak-resistant test suites.
AGNI-Plan
long-horizon task graphs
AGNI-Tools
multi-API orchestration tasks
AGNI-Safety
safe exploration & robustness
Planned cadence: results with every milestone; code when safe.
What We Publish
Research Papers
Peer-reviewed advances in long-horizon intelligence.
Datasets & Benchmarks
Open tasks and evaluation harnesses.
Code & Tools
Open-source implementations where safe and useful.
Research Milestones
Multi-Agent Coordination Framework
Phase 1Negotiation, delegation, and joint planning primitives with evals.
Self-Reflective Architecture v1.0
Phase 2Runtime self-critique, repair, and verification loops for long tasks.
Domain-Specific SLM Suite
Phase 3Compact specialists tuned for targeted workflows and low-latency use.
Unified Evaluation Harness
Phase 4One harness for planning, tool-use, robustness, and safety metrics.
Current Achievements & Vision
Live in production
multi-agent planning prototypes; tool-use orchestrators; early self-reflection loops; internal eval tracks.
Next Phase
unified eval harness; domain-specific SLMs; published benchmarks + ablation studies.
Our commitment to excellence
Rigorous testing through hidden test rotations, adversarial validation, and efficiency-first SLM designs ensure robust, scalable solutions.