Autonomous Coding Agents

7/28/2026

This post details the application of C3 AI Agentic Process Automation (APA) to automate invoice triage within a Sarbanes-Oxley (SOX) compliant environment. It describes the architecture of Invoicing APA, which uses two specialized agents (PO Matching Agent and Invoice Matching Agent) to handle invoice matching and duplicate detection, respectively. Deterministic tools are used for non-judgmental tasks like data pulling and formatting. A key technical contribution is the emphasis on generating a comprehensive audit trail for every transaction, logging the agent's considerations, rulings, and reasoning. The rollout strategy involves starting with a narrow scope and gradually increasing it based on the agent's track record, with a design principle that prioritizes "a miss to a bad post" to minimize audit exposure. The post also highlights the permanent inclusion of a human in the loop for edge cases and exceptions.

7/24/2026

This post introduces the C3 AI Data Intelligence Suite, a new capability built upon the existing Autonomous Coding Agents (ACAs) platform. The suite consists of three specialized agents: Data Interpreter, Schema Creator, and Query Generator. These agents automate the process of data integration by inspecting raw data, building semantic schemas, and generating executable queries based on natural language input. The core technical innovation lies in the agents' execution-grounded approach, where they create, execute, test, and refine artifacts (schemas, scripts, queries) directly against the data within ACAs. This ensures correctness and allows for iterative refinement based on observed failures. The post also details benchmarking results demonstrating the DI agents' performance against state-of-the-art text-to-SQL systems, highlighting the effectiveness of the execution-grounded architecture and schema-aware reasoning.

4/30/2026

Introduced the C3 AI MCP Gateway as an extension to MCP, addressing the limitations of tool-only interaction for enterprise AI agents. The gateway provides a centralized registry for publishing and discovering tools, skills, instructions, and prompts. It implements a curated safety review process for internal skills and external MCP servers. Furthermore, it offers comprehensive observability and analytics across all agent tool usage, including active users, tool/skill usage trends, retention metrics, full trace logging, and audit trails. This enables a feedback loop for optimizing tools and training LLMs, capturing institutional knowledge at scale.

4/30/2026

This post details the technical approach to fine-tuning domain-specific LLMs for autonomous coding agents. It introduces the Narwhal model, a specialist LLM built for the C3 AI Type System. The post outlines the challenges of using frontier models with prompting and retrieval for proprietary enterprise tasks, presenting performance metrics that show Narwhal achieving a 74% pass rate with sub-second latency compared to frontier models at ~60% pass rate and ~2 minutes latency. It details the fine-tuning process involving continued pre-training on proprietary codebases and reinforcement learning (GRPO) with an LLM judge for objective scoring, highlighting how these techniques improve accuracy and generalize across tasks.

3/20/2026

This post details the application of C3 AI's agentic AI platform to supply chain planning. It describes how specialized agents are used to continuously sense change, evaluate trade-offs, and make financially grounded decisions in dynamic supply chain environments. The post highlights the use of technologies like STAFF and multi-hop orchestration agents to connect data from various enterprise systems and create a self-improving network of AI agents for supply chain planning and execution. It contrasts this with legacy systems and provides a scenario of an agentic-driven approach to handling demand surges.

12/24/2025

This post introduces the concept of Autonomous Coding Agents (ACAs) as a new paradigm for enterprise software, moving beyond static deployments to living, self-adapting systems. It details the core components of an ACA (LLM, Coding Agent, Sandbox Infrastructure) and highlights C3 AI's implementation using OpenHands. The post contrasts traditional agentic systems with the new ACA-based paradigm, using the Alchemist system as a case study. Alchemist leverages ACAs to dynamically generate and interconnect optimizer and simulator components, enabling continuous feedback loops and self-healing development through natural-language instructions.

12/2/2025

This post elaborates on the architecture and types of AI agents C3 AI builds for enterprise environments, detailing specialized agents like Retrievers, Code agents, SQL agents, Critics, and Visualization agents. It introduces the concept of multi-agent systems and the trade-offs between single-agent and multi-agent approaches, emphasizing the need for an agentic orchestration layer. The post also highlights the role of the C3 Agentic AI Platform, its model-driven architecture, the C3 AI Type System for data consistency, and the benefits of agent specialization, feedback loops, and modularity for enterprise AI.

11/21/2025

This post details the technical architecture and operational mechanics of agentic AI systems, which are a core component of C3 AI's autonomous coding agents. It elaborates on the concepts of agent roles, orchestration layers, structured reasoning patterns (planning, tool integration, feedback-driven refinement, long-term memory), hybrid agents (combining reasoning and tool execution), dynamic planning, multi-agent coordination, and the practical implementation of these concepts within enterprise workflows. The orchestration architecture is described in detail, including its planning, execution, and integration layers, and how it translates user input into coordinated agentic workflows. The post also touches upon the importance of a unified semantic data model for agent communication and the role of feedback mechanisms for self-correction and adaptation.

11/10/2025

This post details the integration of C3 AI Agentic Process Automation with the C3 Agentic AI Platform, specifically highlighting how the agentic orchestration layer coordinates task-specific agents. It introduces the capability for users to describe problems in natural language, which the system then translates into workflows. The post provides a concrete example of a customer success manager scenario, demonstrating data integration from multiple sources (Google Analytics, HubSpot, Salesforce), ML model execution for churn risk prediction, and subsequent actions to improve account health. It also emphasizes security features like human-in-the-loop review and role-based access controls within the context of automated workflows.

10/31/2025

Introduced Alchemist, an agentic AI system for decision optimization built upon the existing Autonomous Coding Agents framework. Alchemist comprises four specialized agents: Decision Process Extractor, Simulator Creator, Solver Recommender, and Optimization Formulator. These agents leverage LLMs and autonomous coding capabilities to enable subject matter experts to define, simulate, and optimize complex decision systems with reduced reliance on traditional development cycles. The post details the architecture and provides a case study on optimizing nursing staff schedules, demonstrating the agents' ability to handle complex constraints and uncertainties.

10/31/2025

This post details the technical implementation of C3 Generative AI for Code Translation, an extension of the Autonomous Coding Agents capability. It outlines a multi-step process for documenting legacy codebases, including parsing code into abstract syntax trees, mapping dependencies and control flow, identifying external libraries, tracking data flow, and cataloging functional blocks. It also describes the generation of architectural summaries, component descriptions, user stories, API documentation, and explanations of complex algorithms. For code translation, it highlights the use of LLMs within a structured multi-agent framework that incorporates continuous validation, early error detection via compilation and testing, progress mapping, and human-in-the-loop feedback. The post contrasts single-agent vs. multi-agent approaches, emphasizing the benefits of the latter for large-scale code translation.