Agentic AI Platform Change Log

Human approval / guardrails

Version 1.8.0 of Kore.ai's agent platform adds three approval modes to Autoloop: Autopilot applies changes automatically, Copilot applies them only after a person approves, and Advisor only suggests changes. The mode can be set when a loop starts or changed during a run.

Teams can choose how much the automated improvement loop may change their agents without sign off, which matters for regulated deployments that need a review gate before production changes.

Human approval / guardrails

A Human Task node for process workflows arrived in version 1.7.0 of Kore.ai's agent platform. It pauses the workflow, notifies assignees, collects structured input through reusable forms and resumes on the chosen outcome, with deadlines, reminders and escalation.

Teams automating regulated or high value processes get one built in way to put a person in the loop with deadlines and escalation, instead of wiring separate approval steps into every workflow.

Agent capability

Artemis v1.6.5 enables knowledge bases to extract and index mathematical formulas from PDF documents, which previously omitted them. Formula extraction is enabled by default for new knowledge bases, while existing configurations remain unchanged. Gemini Live voice channels also gain controls to enable context compression and configure both its trigger token count and target token count.

Technical document workflows can retain information essential to scientific and engineering answers. Voice teams can tune how long conversations manage their context budget, while existing knowledge bases require a separate decision about adopting the new extraction behavior.

Agent capability

Kore AI for Work version 1.20.0 introduces chat sharing, archiving, project collaborators, and URL-based search alongside improved Slack support. The release also expands model support to include QWEN, adds file generation via Gemini, and splits cryptographic key rotation into distinct DEK retirement and KEK rewrap actions.

Teams can collaborate more effectively on AI projects with new sharing and Slack features, while administrators benefit from broader LLM choices and more granular control over enterprise encryption keys.

Integrations

Kore.ai v11.26.0 added agent transfer summaries that hand conversation context to human agents, expanded prompt context to the last 50 messages, integrated Deepgram Flux for speech recognition, and added Five9 and NICE contact-center integrations.

Transfer summaries and a longer prompt context reduce the context loss that frustrates customers during bot-to-human handoffs, and the Deepgram, Five9, and NICE integrations widen the contact-center stacks the platform plugs into without custom work.

Memory / state

Kore.ai added a personalization layer to AI for Work: Persona, Memory, Custom Instructions, and an Enterprise Search capability. The system builds a persistent per-employee profile that captures role, team structure, aliases, and working preferences, so the assistant carries that context across sessions instead of being re-told each time.

Persistent, per-user memory cuts the repeated context-setting that drags on daily enterprise AI use and pushes responses closer to each employee's actual role and priorities. For buyers, it is a concrete signal of how mature the platform's memory and personalization story is.

Human approval / guardrails

Kore.ai introduced Agent Blueprint Language (ABL), a compiled declarative language for defining and operating enterprise agents. ABL enforces governance, tool usage, and orchestration at runtime through a compiled intermediate representation, treating agent workflows as structured software rather than loose prompt collections.

Deterministic, compiled agent definitions give engineering and compliance teams a handle on prompt-injection risk and behavior drift, and the compiled path also trims multi-agent token consumption.