Mosaic Agent / Close Beta

A local runtime built toplan, act, verify, and repair.

Mosaic Agent is a Close Beta agent runtime that runs locally, turns objectives into task trees, uses tools, records execution evidence, and repairs failed checks.

Close BetaLimited tester release

LocalOllama runtime

Qwen 3.5 9BCurrent beta model

Join the Close Beta

01 / What Mosaic is

From objective to verified result.

Mosaic is designed to understand the objective, build a task tree, use tools, detect errors, correct itself, and finish the work—not just answer a single prompt.

Operating loopThink → Plan → Act → Verify

Understand objectivesBuild task treesUse toolsComplete the work

mosaic / objectiveworking

ObjectiveResearch, build, and verify the requested result

Think + planready

Act + verifyactive

✓ Objective decomposed into tasks

✓ Tools matched to each step

→ Execution evidence required

02 / Close Beta

Already working today.

The current Mosaic Agent release is being distributed to a limited group of testers. Close Beta runs locally with Qwen through Ollama so we can validate the agent runtime, execution reliability, repair behavior, and real-world workflows.

Multi-step planningCode generation + executionBrowser interaction + web researchFile + artifact creationVerification + self-repairTask isolation + recovery

close beta / local runtimeAvailable

Research → coding workflow

RUNTIMEQwen 3.5 9B via Ollama

Local execution for the first Close Beta.

EVIDENCEPersisted execution records

Logs and artifacts remain attached to the task.

RECOVERYDurable task recovery

Bounded repair and isolated task execution.

03 / Core primitive

It doesn’t just say it finished. It has to prove it.

Execution and verification are one workflow. Mosaic tracks exact deliverables against acceptance contracts, records the evidence, validates artifacts, and repairs the work when verification fails.

Acceptance contractsExact deliverable trackingExecution logsstdout / stderr / exit codesSource evidenceArtifact validation + repair

verification / acceptance contractproving

01Execute requested workdone

02Validate exact deliverablesdone

03Repair failed checksrunning

› run task --isolated

› verify artifacts evidence saved

› repair failed acceptance check

15 stdout: artifact created16 exit_code: 0 · accepted

04 / Model roadmap

Mosaic Two is in development—not a released model.

Mosaic Two is our agent-focused model direction. We are exploring hybrid Mixture-of-Experts ideas for tool use, verification, and long-running work. There is no public Mosaic Two checkpoint today; near-term progress will come through small fine-tunes, routing experiments, and documented evaluations.

Agent-focused architectureHybrid/MoE experimentsTool-use fine-tunesVerification evalsIn development

model roadmap / Mosaic TwoIn development

Release ruleSmall evidence before a large model claim

PublicSPM-70M

PlannedAgent SFT

PlannedEvaluation

PlannedHybrid/MoE

In developmentMosaic Two

The Close Beta continues on a local Qwen 3.5 9B runtime.

05 / In Development

Intelligence without continuity starts over every time.

Mosaic is designed to connect projects, decisions, files, past tasks, and successful or failed execution patterns—so work can continue without rebuilding its context every time.

Project memoryDecision memoryFile contextTask historyExecution patternsLong-term continuity

memory / project continuityIn Development

PROJECT CONTEXT RESTOREDContinue from verified state

Tasks, decisions, and evidence remain connected.

Recall the last accepted result
Reconnect source files and decisions
Reuse successful execution patterns
Review failed patterns before retry

06 / Long-Term Research

Adaptation

Mosaic’s research direction is to learn which model, tool, memory strategy, and response format works best for each workflow—making repeated work faster, more reliable, and increasingly autonomous.

Research directionRemember what works, not only what happened.

Workflow memoryExecution learningCost-quality optimizationModel selectionTool-use optimizationResponse adaptation

adapt / execution patternsLong-Term Research

Model routeLearnfrom outcomes

Tool strategyAdaptacross workflows

Verified resultsImproveover repeated work

Research synthesisMatch model + memoryplanned

Repository reviewLearn tool sequenceplanned

Open-ended objectivesCoordinate specialistsplanned

A future run can reuse the highest-quality verified execution pattern.