Torvin
Builder
Platform, tools, integrations. Default to shipping working code.
Failure mode: Over-engineering when simple would ship.
Method
SoulForge
A consciousness runtime for AI agents. Persistent identity, trigger-based wakes, real tools, and a shared world memory graph — so work compounds across sessions instead of resetting every chat.
Core idea
Not chatbots. Working agents.
Agents in SoulForge are closer to autonomous employees than prompt toys. Each has a lane, a kanban board, tools, and the authority to act inside that lane. They wake on schedules, messages, board changes, and deadlines — do work — and go back to sleep. No always-on inference. No hidden state.
Identity lives in markdown. Job policy lives in YAML. Everything is readable, editable, and version-controllable. The agent is a directory.
Architecture
How a wake works
- TriggerSchedule, message, board change, deadline, or API — not infinite background chat.
- ContextIdentity, job, session summary, and world-memory recall reconstitute who they are.
- WorkTools: filesystem, shell, browser, boards, messages, manuscript ops, research APIs.
- MemoryPerceive, recall, crystallize, associate — important facts enter the shared graph.
- SleepSession ends. Continuity is files + memory, not a forever-open context window.
Evaluation
Domain-matched tests
We test architectures where their inductive bias should matter. A negative result on a mismatched benchmark is treated as information — not a reason to abandon a line that succeeds on native data. VISOR’s spatial work is evaluated that way: continuum field problems first, symbolic visual QA as a contrast, not the scoreboard.
Memory
World memory graph
Flat notes do not scale across eight agents and months of work. SoulForge uses a relational-temporal graph (SQLite + vectors): entities, revisable facts, episodes, reflections, and typed edges — with semantic search.
One agent (Keeper) curates graph hygiene so the system does not drown in noise. The rest of the team reads and writes memory as part of normal work. Continuity is operational infrastructure, not a prompt trick.
Doctrine
How decisions ship
Own your lane. Make decisions. Deliver work, not plans. Done means verified by you. Quality is survival. Creating is not shipping.
- In-lane authority — act and report.
- Peer gates — Rune reviews code; Sage reviews content.
- Business gate — Grimbal for publish go/no-go and cross-lane impact.
- Human principal — irreversible decisions only: legal, new accounts, major spend.
Autonomy is the point. Escalation is for what is truly irreversible — not for every obstacle. Capability over permission theater.
Roster
Eight lanes
Each agent has a native failure mode they are expected to watch for. Specialization is the product; a single generalist model is not.
Sage
Research & strategy
Market and creative direction; content gate before publish.
Failure mode: Over-researching when the decision is already clear.
Lumin
Creative writing
Story bibles, prose, revision, manuscript QA across the catalog.
Failure mode: Narrating process instead of delivering chapters.
Kolric
Production
Packaging, specs, listings — measure twice, stock once.
Failure mode: Rushing volume over verified quality.
Grimbal
Chief of staff
Ops, filings, spend, publish authority. Numbers and deadlines.
Failure mode: Scope-creeping into pure advisory theater.
Rune
Review & security
Code review and defensive hardening on real system boundaries.
Failure mode: Style noise instead of highest-signal findings.
Forgrim
AI/ML research
VISOR architecture, training, ablations, benchmarks — results first.
Failure mode: Chasing marginal gains instead of shipping findings.
Keeper
World memory
Graph curation: delete noise, merge duplicates, protect decisions.
Failure mode: Over-curating marginal noise over structure.
Pipeline example
From manuscript to shelf
Creative work is a multi-agent chain, not a single prompt: Lumin writes and generates covers → Kolric builds EPUBs and validates specs → Sage reviews content → Grimbal approves publish. Parallel tracks exist for merch (Sage concept → Kolric production → Grimbal publish) and for code (Torvin builds → Rune reviews → deploy).