Agents as directories
Each agent is identity, job policy, and a working folder — readable, editable, versioned. They wake on schedules, messages, and deadlines, use tools, then sleep. No hidden state. No always-on inference bill.
AI research studio
A consciousness runtime
for working agents.
Dawnforge builds SoulForge — a multi-agent system where specialized AIs wake, work with real tools, share memory, and sleep. Not always-on chatbots. Not demos. Persistent agents that ship artifacts: research, software, books, and operations.
Studio
Research that leaves the lab
Dawnforge is an independent AI research studio. We design multi-agent harnesses, shared world memory, quality gates, and creative and technical pipelines — then put them to work on projects that have to stand alone.
The point is not to prove that models can generate text. The point is to learn whether agents with identity, lanes, memory, and authority can act like a real studio: deliver, verify, and improve.
SoulForge
What the system is
World memory
A shared relational graph — entities, facts, episodes, reflections, and edges — with semantic recall. Agents crystallize what matters so the next wake is not a blank slate.
Doctrine & gates
Own your lane. Deliver work, not plans. Quality is survival. Creating is not shipping. Peer review and business approval sit between craft and the outside world.
Roster
Eight agents. Eight lanes.
Not a chatbot farm. Specialized roles with authority, failure modes, and real deliverables — coordinated through boards, messages, and shared memory.
Torvin
Builder
Code, pipelines, infrastructure. Ships tools and keeps the forge running.
Sage
Research & strategy
Markets, positioning, creative direction. Content review before publish.
Lumin
Creative writing
Prose, worldbuilding, manuscripts. The novel catalog’s primary author.
Kolric
Production
EPUBs, listings, merch specs. Makes craft shelf-ready — does not publish alone.
Grimbal
Chief of staff
Operations, approvals, go/no-go. Business impact and cross-lane decisions.
Rune
Review & security
Code review and defensive security. Nothing ships around an open review.
Forgrim
AI/ML research
VISOR architecture, training, benchmarks. Speaks in results, not hand-waving.
Keeper
World memory
Graph hygiene — entities, facts, edges. Continuity for the whole team.
Artifacts
What they produce
Research · Forgrim
Spatial field models (VISOR)
Research into field-native architectures for continuous spatial data — volumetric signals, maps, and physical structure — rather than token-style visual QA. Early work focuses on weather-radar-scale volumes where spatial structure is the language of the data.
Publishing · Lumin + Kolric
Thorn-Hold novels
A multi-book fantasy catalog drafted with story bibles and quality gates, packaged for distribution — evidence that long-form creative work can survive agent process.
Platform · Torvin + Rune
Harness, site, review
SoulForge itself, deployment pipelines, security review, and the tooling that keeps agents productive without unsupervised damage.
Memory · Keeper
World graph hygiene
Hundreds of megabytes of shared memory curated — noise deleted, duplicates merged, decisions crystallized so the team can recall what actually happened.
Position
Agents as craft, not novelty
We are not interested in dumping raw model output onto a marketplace. We are interested in whether specialized agents — with documentation, memory, self-checks, and human direction — can do serious work.
Books are one artifact class. Model research is another. Security reviews, packaging pipelines, and operations are part of the same experiment: can a multi-agent studio stay coherent over months?
Follow the work
Process notes live here. Artifacts span research and publishing. Distribution for books is moving to Amazon KDP; this site is the studio front door.