Capabilities · Built on Claude
One row per capability, one production surface where it lives, one honest sentence about how. Everything marked Live is running in the product now; Staged means the scaffolding is wired and the optimization is in progress.
Long-context reasoning (≥ 100k tokens per call)
Weekly narrative · /environments/[slug]The 5-sentence Monday memo composes 7 days of AuditLog + signals + goal deltas into one Claude call. One round trip, one memo. Cached 24h.
Sub-agent dispatch (specialist per task)
Project executor chain · /projects/[id]A Project plan fans out to 3–10 specialist executors — Notion fetch, Claude draft, Canva design, Gmail draft. Each gets scoped context. Auto-runs until a human gate.
Tool use with structured schemas
Skill registry · /skill-space14 real tool adapters and a four-dimensional skill classifier (Location, Action, Interaction, Execution). Claude selects from a menu and emits a JSON plan.
Reasoning-trace exposure
WhyDrawer on every action · /environments/[slug]Click any Atrium action in the ledger. Drawer shows what Atrium read, what tool it called, what it produced, tokens and cost. KernelTrace + IntelligenceLog, no black boxes.
Persistent memory per organization
Atrium Memory + Mastery Insights · /memoryAtriumMemory holds user-taught corrections; MasteryInsight holds patterns Atrium derived from execution reviews; OperationalPlaybook holds the regenerated "how work actually runs" doc. All org-scoped.
Bidirectional teaching loop (user ↔ model)
Atrium Academy · /learnFour fluency capabilities (delegation, review, context-giving, trust-calibration) scored from aggregates. Daily lesson biased to the weakest. Every answer writes a AtriumMemory entry future Atrium calls RAG against.
Streaming build narration (SSE)
Onboarding build step · /welcomeAtrium streams scaffold progress as it creates the System and Workflow. Real steps, real writes, rendered one line at a time.
Reversible autonomous actions
24-hour undo window · /environments/[slug]Every autonomous action gets an AuditLog undo path and a AtriumMemory user_correction on reversal. Future Atrium calls factor in the correction.
Calibrated autonomy (5 levels)
AutonomyConfig per Workflow · /approvalsObserve → Suggest → Act & Notify → Autonomous → Self-Direct, per Workflow and per System. Trust Score recalibrates from approval rate.
Override reason capture (structured)
Teach Atrium pill row · /environments/[slug]Three override categories plus one-line context, per rejected action. Written to AtriumMemory as type=user_correction with category set.
Prompt caching (cost shape)
Narrative + planner calls · server-sideSystem prompts are structured so the static prefix is cache-friendly. Cache adoption is a continuing optimization, flagged here so partners know we track it.
MCP-compatible tool contract
Skill adapter shape · lib/skills/executors/Every tool adapter implements a single Executor interface (step, project) → ExecutorResult with artifacts, trace, mode. Maps cleanly to MCP when we expose it externally.
The architecture brief that expands every row lives at /security/architecture. The roadmap for what comes next lives at /roadmap.