Mission
Lucidic AI is building the training and optimization layer for AI agents, so enterprises can measure, improve, and safely deploy agents that actually work.
We build tooling that automatically evaluates and improves enterprise AI agents using search and learning loops across prompts, tools, and policies. Our customers care about reliability, auditability, and measurable gains.
Our approach
People get better in more than one way: they build skill, and they get better instructions and tools. Agents are the same. Improving an agent means improving the model and the harness around it.
- Harness optimization
- Prompts, tools, context, memory and policies turn into tunable parameters. The model weights stay frozen. Lucidic inserts training modules where your agent fails, tunes them, and tests every candidate in offline simulation.
- Model-harness co-training
- Your harness stays as it is, with every parameter still tunable. Now the model weights are on the table too, and the two are trained as one system. Lucidic sorts your traces and sends every case to the method that learns from it best: LoRA SFT, on-policy distillation, training modules or GRPO.
- Your data
- Company and production data become datasets and learnings that keep improving your agent over time. Your agent produces thousands of traces, and most are routine. We group similar ones and keep what matters: unusual cases, failures, and one example of each common pattern. The result is a small dataset that shows how your agent really behaves, at a fraction of the cost.
- Simulation
- We build a replica of your agent's world: its tools, its databases and simulated customers. The agent runs through it step by step, and when a run goes wrong, we go back to the exact step where it slipped. From there we replay with a different choice, again and again, keeping only the paths that score higher, until it reaches the reward.
- New models
- When a better model comes out, you do not start over: your datasets and learnings carry over to the new model.
Results
-
65%vs 51.1% for the best frontier model
Strict task success when ServiceNow evaluated Lucidic's training method on 1,150 tasks drawn from real enterprise workflows (EnterpriseOps-Gym). ServiceNow case study
-
24 hoursvs months of manual work
Cresta's FECT contact-center benchmark went from 59% to 87.5% accuracy in under 24 hours, for about $200 of API spend instead of more than $100k. Cresta case study
-
92%up from a 75% plateau
Accuracy of an enterprise customer-support agent once Lucidic optimized all 116 parameters across it and its sub-agents, after 2-3 months of manual tuning had stalled at 75%. Enterprise case study
-
92.3%vs 87.9% for the next best
On Sierra's τ²-bench for customer-support agents, ahead of Qwen3.5-397B (87.9%), Claude Opus 4.5 (85.3%) and GPT-5.2 (84.8%). Benchmark results
Public benchmarks
| Benchmark | Lucidic | Next best |
|---|---|---|
| Harvey’s LAB benchmark | 24% | 19.7%, Harvey Tenet (fine-tuned Kimi K3) |
| Sierra’s τ²-bench | 92.3% | 87.9%, Qwen3.5-397B (open weights) |
| Mercor’s APEX-Agents | 52% | 43.5%, Claude Opus 5 |
| Zapier’s AutomationBench | 70% | 52%, Gemini 3.7 Flash |
| HotpotQA · exact match | 44.5% | 4.7%, DSPy GEPA |
| HotpotQA · F1 | 0.7 | 0.66, DSPy GEPA |
| IFBench | 56.4% | 51.2%, DSPy GEPA |
| PAPILLON (PUPA) | 98.3% | 94.47%, DSPy GEPA |
Methods, sources and every entry: Results on the homepage.
Research
Security
Lucidic is SOC 2 Type II and HIPAA compliant. Our Trust Center has our security documentation and how to request our reports. For anything else about security, email team@lucidic.ai.
Careers
If you want to ship agentic systems that get better with real feedback loops, come build with us in San Francisco.
- Founding Applied Research Engineer San Francisco, CA · On-site
- Member of Technical Staff San Francisco, CA · On-site
- Research Systems Engineer San Francisco, CA · On-site
- Infrastructure Engineer, ML Systems San Francisco, CA · On-site