Mosaic Agent / Core capabilities

Intelligence thatremembers the work.

Plan, execute, remember, and improve—while using only the intelligence and context each task actually needs.

06Core capabilities

PersistentProject memory

AdaptiveModel routing

Explore capabilities

01 / Capability

Minimal Token Consumption

Mosaic Agent completes tasks with the lowest practical token usage. It removes unnecessary context, compresses long conversations, retrieves only relevant memory, and sends simple work to smaller models first.

Core ideaEvery token should have a purpose.
Relevant context onlyConversation compressionSmaller-model routingReasoning on demand
context / optimizer84% saved
Current taskCompare three research papers
Full context32.4K
Relevant memory2.18K

Compressed 18 conversation turns

Retrieved 4 relevant memories

Routed to Mosaic Core

02 / Capability

Memory-Native Work

Mosaic Agent remembers users, projects, preferences, decisions, and previous work across sessions—without repeatedly sending the complete conversation history.

User memoryProject memoryTask memoryDecision memoryWorkspace memoryRelevant retrieval
memory / project-atlas4 relevant
Continue the launch plan from Tuesday.
DECISIONLaunch sequence approved

Use the staged rollout with a 10% first cohort.

PROJECTAtlas milestone 04

Evaluation suite is complete; documentation remains.

PREFERENCEConcise weekly updates

Report only blockers, decisions, and next actions.

03 / Capability

Planning and Tool Use

Mosaic Agent understands the objective, creates a plan, selects the right tools, executes the work, and verifies the result. It is designed to act, not only respond.

Task planningStep-by-step executionResearch workflowsCode executionFile analysisResult verification
workspace / releaseexecuting

01Inspect repositorydone

02Find failing routedone

03Patch and verifyrunning

grep "rateLimit" src/

read_file src/api/routes.ts 42 lines

edit src/api/routes.ts

15 await rateLimit(req, { max: 100 });16 return schema.parse(data);

04 / Capability

Intelligent Model Routing

The largest model is not used for every request. Mosaic Agent selects the smallest capable intelligence layer based on difficulty, cost, context, and required reasoning depth.

Mosaic CoreMosaic MemoryMosaic ToolsMosaic TwoAgent Runtime
router / intelligenceoptimized
Incoming taskReview a complex architecture decision
SimpleMosaic Core
MemoryCore + Memory
ActionsMosaic Tools
ReasoningMosaic Two
Long projectAgent Runtime

Selected for reasoning depth and verification requirements

05 / Capability

Workspace Continuity

Mosaic Agent is organized around persistent project workspaces rather than isolated chats. Files, decisions, plans, and task history remain connected to the work.

Project-based memoryLong-running tasksFile associationDecision trackingProgress awarenessMulti-session continuation
projects / mosaic-runtimesession 12
RESUMED FROM LAST SESSIONDistributed evaluation pipeline

3 of 5 tasks complete · 2 decisions retained

  • Define evaluation inputs
  • Implement result collector
  • Verify failure recovery
  • Document operating limits

06 / Capability

Mosaic Adapt

Mosaic Adapt learns which model, tool, memory strategy, and response format worked best for each workflow—making repeated work faster, cheaper, and more accurate.

Core ideaMosaic Agent does not only remember information. It remembers what works.
Workflow memoryExecution learningCost-quality optimizationModel selection improvementTool-use optimizationResponse adaptation
adapt / workflow intelligencelearning
Token cost−31%last 30 runs
Latency−24%optimized route
Success96%verified output

Research synthesisMosaic Core + Memorybest route

Repository reviewPlanner + Tools−18% time

Weekly reportingSaved response stylepreferred

Next run will reuse the highest-quality execution pattern.

Mosaic Agent

Do more work.
Carry less context.

Follow development