Agents
Most writing tools that ship a chatbot give you a single one. You type, it responds, and whatever it thinks about your question is colored by the fact that it’s one general-purpose assistant trying to be good at everything. The output is mediocre in three directions at once — a generic writing voice, a generic research voice, a generic mechanics voice — none of which is actually what you wanted. Ishvana takes a different approach. Instead of one chatbot, there are six agents, each with a distinct domain, distinct tools, distinct memory, and a distinct job. You talk to the one whose job matches what you need, and the output is shaped by a specialist.
The six agents are all powered by the Divinity Engine — bundled deterministic handlers running in your local backend, no external service contacted at runtime. None of them generate prose by sampling tokens from a foundation model; they read your work, run typed analyses against it, and surface findings through pre-written prose templates the Authored Library keeps deterministic across runs.
The roster splits into two groups. Divine agents are the top-tier specialists with chat surfaces and the broadest tool sets. Standard agents run focused analysis tasks and don’t carry on open-ended conversations.
Divine agents
Section titled “Divine agents”Standard agents
Section titled “Standard agents”How to know which agent to talk to
Section titled “How to know which agent to talk to”The simplest answer: talk to Ishvana first. Her intent classifier figures out which specialist should handle a question and routes the request there.
If you already know what you need, talk to any agent directly from the Chat tab. Pick the one that matches:
| Your question is about… | The agent to talk to |
|---|---|
| Your world, your characters, your canon | Ishvana (or she’ll route as needed) |
| Prose feedback, style metrics, scene-level reading | Hawken |
| Research, web content, monitored sources | Lagan |
| Stat blocks, abilities, encounter balance, system mechanics | GameMaster |
| Real-world facts, fact-checking prose, world rules | WorldKnowledge |
| Consistency checking against your established canon | Lorekeeper |
These aren’t rigid walls. Hawken can answer basic lore questions without bouncing you to Ishvana; Ishvana can summarize a research finding without asking Lagan; Lagan can quote a Wikipedia article without involving WorldKnowledge. The specialization is about who is best at the job. Anyone can touch anything if they have the handlers for it.
What each agent runs on
Section titled “What each agent runs on”Every agent maps to a set of Divinity Engine handlers documented in Etherforce → Handler registry. When you ask Hawken to read a scene, you’re triggering divinity/hawken-scene-scan, which composes ProseGuard + Worldspell + Insights + Craft Plot + scene contradictions + sentiment + register + scene metrics. Every handler is local, deterministic, and reproducible.
This is different from how older Ishvana versions worked. Pre-2026-04, the agents routed to cloud LLM providers (Anthropic, OpenRouter, Ollama) and the output varied between runs. The Divinity Engine replaced that layer with bundled handlers, and the upshot is reproducibility — same scene plus same lore plus same parameters always produces the same findings.
Where each agent lives
Section titled “Where each agent lives”Divine agents have a home module where they’re most useful, and they’re also reachable from the Chat tab.
- Ishvana lives in Chat as the default agent. She’s everywhere.
- Hawken lives in the Writing module’s Hawken panel and slash commands. She’s the agent you reach for while reading what you wrote.
- Lagan lives in the Research module — the Lagan Chat tab, the Pipeline tab, and the Monitored Feeds tab.
- GameMaster lives in the Magic System workspace as a specialist panel, reachable from characters and stat blocks.
- WorldKnowledge lives in the Chat tab and is also tied to the WorldKnowledge subtab in the Legendry where world rules and Wikipedia search are managed.
Standard agents:
- Lorekeeper lives in a dedicated panel accessible from the Writing module and the Analysis workspace. He doesn’t have a chat mode — you ask him for a consistency check, he runs it, you review the results.
Parallel composition
Section titled “Parallel composition”When a request touches multiple domains, the routing dispatches multiple handlers in parallel and Ishvana assembles the joint result. “Research the political structure of this faction AND check whether the diplomatic scene I drafted holds together” fans out to Lagan’s research handlers and Hawken’s scene scan at the same time, then synthesizes the findings. This isn’t a UI mode you opt into — it’s something the orchestrator does when a request warrants it. Complex multi-domain requests resolve faster than they would with a one-agent-at-a-time chain.
Agent transparency
Section titled “Agent transparency”Every handler dispatch is tracked — which agent triggered it, which handler answered, the applied parameters, the duration, the success or failure. You can see this data in the Agent Overview panel in the Analysis workspace and per-dispatch in Etherforce Observability. This exists because you deserve to see how well the engine is doing on your specific project, not just trust that everything is working.