Salesforce drops Koa enterprise agent model

// 1h agoRESEARCH PAPER

Salesforce drops Koa enterprise agent model

Salesforce published research detailing Koa, a 120-billion-parameter enterprise language model optimized for agentic tool use and multi-step CRM workflows within its Agentforce platform. Built by post-training NVIDIA's Nemotron-3-Super-120B with GRPO, Koa uses a simulation-to-reward pipeline based on declarative Agent Script configurations to match frontier models on enterprise benchmarks without customer data.

// ANALYSIS

Training enterprise models directly from the declarative configuration scripts used to deploy agents is the blueprint enterprise AI has been waiting for.

–Specification-driven RL: Using declarative Agent Script files to generate synthetic simulation tasks turns software architecture configurations into reinforcement learning flywheels.
–Domain specialization beats frontier brute force: A targeted 120B open-weight foundation model tuned on workflow topologies outperforms generic, expensive mega-models on complex CRM automation.
–Zero customer data risk: Relying entirely on synthetic tasks generated from schema definitions and domain priors bypasses enterprise compliance and data privacy bottlenecks.
–Salesforce's true moat: Owning the orchestration DSL, schema definitions, and workflow execution runtime matters far more than training foundation models from scratch.

// TAGS

salesforcekoanemotronagentforceagentsenterprise-aigrporeinforcement-learningllm

DISCOVERED

1h ago

2026-09-16

PUBLISHED

1h ago

2026-09-16

RELEVANCE

8/ 10

AUTHOR

omarsar0

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

NEWS1h ago

OpenAI tops Anthropic in OpenRouter wallet share

According to recent weekly data from LLM gateway OpenRouter, Anthropic has lost its majority developer wallet share for the first time in over 133 weeks, with OpenAI overtaking it in total dollar spend. Because wallet share tracks actual credit card and enterprise billing rather than raw token volume, the transition underscores how quickly production workloads migrate in a market where unified routing platforms eliminate switching friction.

INFRA2h ago

Abnormal AI deploys Amazon Bedrock AgentCore at scale

Abnormal AI has integrated Amazon Bedrock AgentCore into production to conduct inline, real-time email threat detection across billions of messages. By leveraging AgentCore's serverless runtime and secure Code Interpreter inside ephemeral microVMs, Abnormal's autonomous security agents dynamically execute code and programmatically verify complex cybersecurity threats before they hit user inboxes.

BENCHMARK2h ago

Agent Memory Leaderboard drops Coding Memory benchmarks

The Agent Memory Leaderboard (AML), an open benchmarking initiative co-organized by international academic and research institutions, has released its initial evaluation results for its Coding Memory track. Designed to assess how AI agents store, retrieve, and operationalize software engineering experience across multi-session workflows, the benchmark evaluates both commercial products and open-source methods across 12 repositories and over 1,200 historical tasks.