Kore.ai news and updates

Kore.ai's Artemis agent platform v1.8.2 lets builders list the exact session fields the model may write in a reasoning step, instead of only turning model writes on or off. An empty list blocks all model supplied updates in that step, while authored actions such as SET, GATHER extraction, hooks and tool result mappings work as before. Existing agents keep their current behavior until a step is changed.

Our read

Teams building servicing agents can stop the model from writing fields such as verification flags or amounts, leaving those to authored logic. Every Runtime instance must run v1.8.2 before an agent that uses field lists is published.

Kore.ai Agent Platform 1.8.0 gives Autoloop, its evaluation driven loop that proposes improvements to agents, three approval modes. Autopilot applies changes automatically, Copilot applies them only after a person approves, and Advisor only suggests them. The mode can be set when a loop starts or changed mid run, and it carries over when a loop restarts.

Our read

A bank can hold agent optimization to Copilot or Advisor so no AI proposed change reaches a live agent without a named reviewer, which fits model change control. Autopilot is now also on the menu, so governance teams should decide which modes they allow.

Version 1.7.3 of Kore.ai's Agent Platform changes how agents handle several requests in one message: two new settings decide when a request counts as finished and have the supervisor capture every request at once, and the automatic strategy now handles requests one after another and asks the customer when they are ambiguous. When a tool call needs approval, customers now see a plain question built from readable labels such as account and amount, with sensitive values masked and internal instructions hidden. Tool hooks can pass the model a short summary instead of a large result, and project imports keep their model selections.

Our read

Banks running customer facing agents get clearer consent prompts before an agent moves money or changes an account, and fewer dropped or misordered requests. Teams that relied on parallel handling of requests now have to set it explicitly.

Version 1.7.2 of Kore.ai's Agent Platform lets builders choose how much of each tool result the model keeps in conversation history (the full result, selected fields or a template summary) and treats workflow and MCP tools as having side effects by default, so retries do not repeat operations and confirmation policies cover them. It adds an Oracle Database connector for queries and row changes, optional encryption of payloads returned by the Sessions, Conversations, LLM Logs and Traces APIs, filtered bulk delete for Agent Tables up to 10,000 rows, and an FAQ connector for knowledge bases. The idempotency header on HTTP tools is now opt in, and the voice guardrails introduced in 1.7.0 are now off by default.

Our read

Banks building agents on Kore.ai get protection against duplicate transactions on retries, encryption for exported conversation data and a direct connection to Oracle databases. Teams that relied on the 1.7.0 voice guardrails need to switch them back on for each agent.

Kore.ai AI for Work v1.21.0 adds an Enable guardrail toggle to the Tools section of an MCP server, so account guardrails run before any tool executes. The release also gives administrators a Models tab in Workspace Settings to enable, disable and reorder the LLM models available to each workspace.

Our read

Banks and insurers letting employee agents call tools over MCP get a pre execution control point and can restrict which models each team uses, which supports model inventory and human oversight expectations.

Kore.ai released Agent Platform v1.6.3, which adds a command line tool for automating the agent lifecycle, support for the Microsoft Foundry model catalog, and PII detection by entity. It also routes all outbound traffic through configurable proxies and adds evaluation suites for each environment.

Our read

Proxy control of every outbound call and entity level PII detection are the items that usually hold up a bank's security review of an agent platform. Evaluation suites per environment let teams run the same checks in staging that they will run in production.