Co-training harnesses and models together
Frontier Intelligence, Shaped Around Your Data.
We research the psychology of how LLMs learn, and turn it into new training techniques that train custom models and harnesses together.
AI Agent
System Prompt
You are a helpful AI assistant trained to provide accurate and thoughtful responses. You approach each question with care, considering context and nuance. Your goal ...
Tool Call Descriptions
search_web(query="latest AI research") read_file(path="/data/context.txt") write_file(path="/output/result.json", content="{...}") execute_code(language="python", code="import numpy...") call_api(endpoint="https://api.example.com/v1/...", method="POST")
Tools
Search
Database
Code
Tools
Guardrails
CRITICAL SAFETY RULES: - Never execute destructive commands without confirmation - Validate all user inputs for SQL injection patterns - Rate limit: max 100 requests per minute - No access to production databases - Require authentication for sensitive operations - Log all security-relevant events
Policies
BEHAVIOR POLICIES: - Maintain professional, helpful tone - Cite sources when providing factual information - Admit uncertainty rather than guess - Respect user privacy - no data retention - Follow ethical AI guidelines - Provide explanations for complex decisions
Context
CURRENT CONTEXT: User: Data Scientist Task: Analyze sales trends Dataset: Q4_2024_sales.csv Environment: Production Timestamp: 2024-01-16T14:30:00Z Session ID: a8f3d2c1-b4e5-4f6a-9d7e-1c2b3a4d5e6f Previous queries: 3
Memory
SHORT-TERM MEMORY: - User asked about sales trends 2 minutes ago - Loaded dataset: 45,000 rows - Identified top product: Widget Pro X - Found anomaly in December data LONG-TERM MEMORY: - User preferences: Python over R - Common visualizations: bar charts - Typical workflow: clean → analyze → visualize
Models
GPT-4
Claude
Gemini
Llama
Initializing Training...
Trusted and supported by engineers at
Our Philosophy
Training a model is like teaching a human.
People learn in different ways: taking notes, solving practice problems, using flashcards. Good teaching starts with understanding how the learner thinks: where they get stuck and what kind of feedback changes their behavior.
We research those same questions in LLMs to train models and harnesses together through adaptive simulations, reflective evolution, inference-time optimization, and more. We've seen this combined approach significantly outperform both weight-only post-training and harness-only optimization at a fraction of the cost & data. If you care about building the best agents, we'd love to talk.
How It Works
PHASE 1
Define the Training Surface
Start by choosing which parts of your agent are trainable. That includes the harness — system prompts, memory, context, tools, guardrails, policies — and the model itself, treated as a trainable policy rather than a fixed dependency.
Training both together costs a fraction of weight-only post-training — harness changes take no gradient steps and are optimized through a custom algorithm and reward function.
Parameters
Results
Logs
Agent Parameters
Auto-tune
Memory
System Prompt
Models
Context
Tools
Guardrails
PHASE 2
Explore the Behavior Space
Next, turn your production traces and datasets into environments. Lucidic runs rollouts across them, varying the harness and the model's policy systematically to learn which choices earn rewards on your cases and which don't.
Lucidic's simulation builder turns your production data into synthetic environments, with your tools and data mocked. Rollouts run fast, and any production trajectory can be replayed or branched from any step against a new harness or model.
Simulations
History
Settings
Decision Tree
Exploring optimal paths
Live
PHASE 3
Co-train Harness and Weights
Finally, train both halves on the same signal. Lucidic uses genetic algorithms to search over harness configurations while running RL policy optimization on the highest leverage training tasks, so each improves in step with the other.
At inference, the trained harness runs alongside the model with tools it calls — consulting a knowledge module, verifying a write before it commits, recalling a past successful run — all loaded from one checkpoint through the SDK.
Metrics
Charts
Export
Performance Metrics
Real-time analytics
Auto-refresh
Overview
Accuracy
Variance
Iterations
Support
Deflect Rate
Resolve Rate
Fake Deflect
Better Than State of the Art
Real-World Applications
Sierra: Customer Support Agents
Sierra builds customer-service agents for large consumer brands. Its τ²-bench puts an agent on an airline, retail, or telecom support desk. The agent has to follow the company’s written policy, look up and change records with tools, and talk a simulated customer through the fix. A task passes only if the final account state matches what the policy says it should be.
Lucidic AI
92.3%
Qwen3.5-397B
open weights
87.9%
Claude Opus 4.5
85.3%
GPT-5.2
84.8%
Gemini 3 Flash
83.5%
τ²-bench overall pass^1, the mean of the retail, airline, and telecom domains. Sierra-verified leaderboard runs on tau2-bench v1.0.1 with the GPT-5.2 user simulator. Sierra has not verified a three-domain overall for newer models.
Case Studies
Real results from agentic training
Customer satisfaction rate was a crucial metric that was dangerously low, but after using Lucidic's auto-improvement algorithms, it improved significantly.
6 months
Time saved by using Lucidic's auto-improvement algorithms
48%
Relative improvement from customer resolution rate baseline
Cresta Case Study
Read morePlatform Features
Everything you need to train your agentic systems
Works with your stack
Agent Integration
Integrate with any LLM provider and agent framework. LangChain, LangGraph, Langfuse, OpenAI, Anthropic — Lucidic works with what you already use.
Define success
Custom Reward Definition
Define objective functions aligned with domain-specific metrics — inference latency, computational cost, and other measurable outcomes.
Explore possibilities
Intelligent Candidate Exploration
Automatically search thousands of agent configurations — prompt variants, tool orderings, context strategies — to find what works best.
Ship safely
Continuous Improvement in Production
Deploy improved agents with controlled rollouts. Gradually shift traffic, monitor for regressions, auto-promote or rollback.
Dive into Research
Explore Our Latest Research
Discover insights from our ongoing research into AI agent optimization, performance analysis, and evaluation methodologies.
View Research
Applications of Training
We help teams solve the hardest challenges in agentic deployment
Train against outcomes that matter
Optimize Your Metrics
Train agents on what matters in production: accuracy, CSAT, resolution rate, escalation rate, and more. Lucidic optimizes toward the metrics that define success for your product.
Frequently Asked Questions
Who should use Lucidic?
Lucidic is for teams that want more reliable, higher-performing AI agents. Whether you fine-tune models or rely on off-the-shelf LLMs, Lucidic helps teams systematically all parameters of the agent without manual iteration.
Does Lucidic AI build agents?
No. We don’t replace your agent architecture or infrastructure. We optimize and train the agents you already have, focusing on reliability, accuracy, and continuous improvement.
How does the training process work?
Lucidic runs structured simulations that systematically vary agent design choices across scenarios. Performance is evaluated against your chosen metrics, and learning algorithms—including evolutionary methods, reinforcement learning, and Bayesian optimization—use this signal to iteratively improve agent behavior over time.
Do I need to rebuild my agent to use Lucidic?
No, Lucidic integrates with all major LLM providers and agent frameworks. We don’t rebuild or replace your agents—instead, we train and improve the components you already use, including prompts, tools, policies, and decision logic.
What types of AI agents can be trained?
Lucidic works with any LLM-powered agent including customer support agents, coding assistants, data analysis tools, and custom enterprise agents. If it uses an LLM, we can optimize it.
Do you need a training environment and reward to use Lucidic?
No—Lucidic can work with what you already have. If a training environment or reward signal isn’t fully defined, our team can help design and refine them as part of the engagement. Clearer environments and richer signals enable stronger learning, and we work with customers to progressively build those pieces over time.
