Outcome

Agents that write production-quality code

Without context, agents generate implementations that break patterns, miss dependencies, and fail tests.

Unblocked gives them the architectural decisions and conventions behind your system so the first draft actually fits.

Without Unblocked

Broke backwards compatibility

12 compilation failures

Cascading module breaks

Quality Bar

FAILED

With Unblocked

Backwards compatibility preserved

Tests pass across modules

Matches team patterns

Quality Bar

PASSED

The experiment

We ran the same engineering task twice: once with Unblocked and once without it, using the same coding agent, codebase, and prompt.

Task: Introduce adaptive thinking mode for supported Claude 4.6 models while maintaining backwards compatibility with existing code using the deprecated budget_token method.

The results with Unblocked

Plan

1.Auto-detect adaptive mode for 4.6 models

2.Convert ReasoningEffort to auto-switch

3.Add custom serializers for null handling

4.Add unit tests for thinking mode

Outcome

Broke backwards compat (auto-detect, no opt-in)

12 compilation failures across modules

Cascading Bedrock module breaks

Batch provider API redesigned 3 times

Unnecessary custom serializers added

Quality Bar Failed

Plan

1.Opt-in via useAdaptiveThinking: Boolean

2.Model-aware config (supportsAdaptive flag)

3.All modules: direct API, Bedrock, batch

4.19 unit tests + API integration tests

5.Nullable fields via kotlinx defaults

Outcome

Opt-in preserves all existing behavior

Caught batch provider gap in plan review

No custom serializers needed (simpler)

Extension functions match team patterns

Quality Bar Passed

Validation & Integration Tests

Without UB:

2.5/10

With UB:

9/10

Without Unblocked:

Unit tests only — no API integration, no Bedrock or batch coverage. Bugs in 2 of 3 modules would ship undetected.

With Unblocked:

19 unit tests + API integration tests across all 3 modules. Regressions caught before merge.

Doesn't Break Existing Implementation

Without UB:

2/10

With UB:

9/10

Without Unblocked:

Auto-detect silently changed behavior for all budget_token callers — breaking every existing integration if not caught by human review.

With Unblocked:

Opt-in default (false) preserves all existing behavior. Zero risk to current callers, no review flags.

DRY Code, High Quality Bar

Without UB:

3.5/10

With UB:

9/10

Without Unblocked:

Added custom serializers the team doesn't use, redesigned the batch API 3 times. All of it would need to be ripped out in review.

With Unblocked:

Used kotlinx defaults and existing serialization patterns. Single clean pass, nothing to rework.

Respects Our Conventions

Without UB:

2/10

With UB:

9/10

Without Unblocked:

Used a factory method pattern foreign to the codebase, causing 12 compilation failures. Every module would need rewriting to match team style.

With Unblocked:

Extension functions match existing team patterns. Zero compilation failures — merged without style feedback.

Agents that write production-quality code

The experiment

The results with Unblocked

With context, the agent passed every quality check. Without it, the output cost more to correct than it saved to produce.

Quality scorecard

Save tokens on every task