Standard HarnessIndependent systems lab Singapore

Agent systems that finish.

Standard Harness improves verified completion while reducing cost, latency, and unnecessary context. The Singapore-based lab behind Codegraff builds repository intelligence, model routing, tools, sandboxes, and verification for coding agents.

See how it works Explore Codegraff
SH–000 / Field image
Figure 00.1Independent systems lab
DONECompletion

Did the work actually finish?

COSTEconomy

What did the result consume?

TIMEPerformance

How quickly and reliably did it run?

ENVInfrastructure

Which context, model, tools, and sandbox?

Task to verified result.

One operating loop surrounds the agent: map the work, choose a route, execute inside a boundary, verify the result, and keep what improves the next run.

An inked turtle engineer turns repository task cards into verified result cards through a six-channel harness machine.
One loop / six checks
  1. 01Workload
  2. 02Context
  3. 03Route
  4. 04Execute
  5. 05Verify
  6. 06Improve

Four jobs, one operating layer.

Each part solves a different source of waste between a repository task and a verified result.

01 / ContextMap only what matters.

Definitions, callers, outlines, and dependencies replace repeated full-file reads.

A bespectacled turtle turns scattered repository cards into one structural map.
02 / RoutingChoose the right path.

Match the task to a model and strategy by completion, cost, latency, and limits.

The turtle selects one of three model and strategy routes.
03 / ExecutionRun inside a boundary.

Metered sandboxes isolate commands, preserve logs, enforce limits, and verify the work.

The turtle monitors a task running inside a contained, metered chamber.
04 / LearningKeep what earns it.

Verified outcomes improve future strategies without collapsing every task into one winner.

The turtle archives verified results and returns them to the next work cycle.

How we judge improvement

Completion is the gate. Everything else is optimized around it.

A cheaper or faster run is only better when it preserves correctness, final state, and reliability under the same workload.

SH / Measurement consoleOutcome resolved
Primary gateVerified completion
CompletionCorrectness gate
CostOptimize
LatencyReduce
ContextFocus
ReliabilityPreserve

Research becomes working infrastructure.

Four systems carry the ideas into repository understanding, agent strategy, local execution, and model access.

A

Evolves

Graff

Runs strategy variants on real work, scores what happened, and reuses what earns it.
B

Maps

CodeDB

Builds a local structural map of symbols, callers, outlines, and dependencies.
C

Moves

CodeDB Pro

Keeps local tools warm, batches operations, and reduces unnecessary context and waiting.
D

Routes

Gateway

Connects model access, cost metering, limits, and sandbox execution behind one boundary.

Standard Harness / Singapore

Understand the work. Run it well. Prove the result.

Explore Codegraff Contact the lab