Does Debrix replace my observability platform?

No. Keep your existing platform for production monitoring, detection, clustering, and impact measurement. Debrix focuses on the controlled local investigation of failures.

How does controlled debugging work?

Debrix freezes the original evidence, changes one event or input, and records which later steps were recorded, edited, or live. Comparing the branch with the original helps narrow the fault boundary.

Does a passing diagnostic branch mean the failure is fixed?

No. A passing branch is evidence about where to investigate. Debrix marks a Failure fixed only after the real code, prompt, schema, model, or policy change passes the original scenario without exploratory overrides.

How does Debrix know whether a run succeeded?

You supply or approve the outcome contract: a test, assertion, state predicate, human verdict, or clearly labelled probabilistic judge. Debrix records pass, fail, or inconclusive together with that validation source.

Does Debrix require a custom agent runtime?

No. Your existing agent and codebase continue to run normally. Debrix owns the experiment plan and provenance, while you or your coding agent run the project through the SDK, MCP, or an approved local workflow.

Can I download Debrix today?

There is no public desktop download yet. The macOS app is distributed in small early-access batches; the Python SDK and IDE skills are open source. Selected engineers get their first three months of Pro free.

Debrix · local-first macOS debugger

The interactive debugger for AI agents

Open a failed run. Hold the evidence constant. Test one change at a time. Then verify the real fix with no overrides and save it as a regression.

Selected engineers get their first three months of Pro free.

See the failure investigation

Failure investigation

One failed run, held still

destinations.length >= 1

Original run

Failed

Recorded evidence

search_places({ query: "France" }) → []

Branch B1

Diagnostic

Override result

Downstream path passes

Branch B2

3 / 3

Correct arguments

Tool boundary narrows

Verified fix

No overrides

Require country_code

Regression

Passed

destination_suggestions_not_empty

Product preview · Failure investigation

One failure. Multiple controlled branches. One verified fix.

Select an experiment to see what stayed recorded, what changed, what ran live, and what the evidence actually proves.

A diagnostic pass narrows the fault boundary. Only the production change, rerun without overrides, verifies the fix.

Empty destination suggestions

Failure #42 · travel-agent

Investigating

Outcome contract destinations.length >= 1

Investigation

Regression

destination_suggestions_not_empty

Passing

Original

Freeze the failed execution

Failed

Starting point

Capture the completed run and its violated expectation without assigning a root cause.

Prompt

Plan a weekend in France.

Recorded

Tool call

search_places({ query: "France" })

Recorded

Tool result

[]

Recorded

Final response

No destinations found.

Recorded

Execution provenance

Prompt: Recorded
Tool arguments: Recorded
Tool result: Recorded
Final response: Recorded
Outcome contract: Failed deterministic assertion

Expected: ≥ 1 destination
Observed: 0 destinations

Bounded finding

The run completed, but its outcome violated the product expectation. The evidence says what happened—not why.

Next focus: Reproduce this baseline, then change one boundary at a time.

Branch B1

Test the downstream path

Diagnostic pass · not fixed

Hypothesis

The downstream planner may work if it receives valid destination data.

Tool result

Recorded

[]

Temporary override

["Paris"]

Final response

Original

No destinations found.

Live continuation

Paris itinerary ready.

Execution provenance

Prompt: Recorded
Tool arguments: Recorded
Tool result: Edited
Final response: Live
Outcome contract: Passed deterministic assertion

Original: Failed
Branch B1: Passed

Bounded finding

The downstream planner can produce a valid itinerary when it receives destination data.

Next focus: Test whether the real tool can return valid data when given a stricter argument.

Branch B2

Test the tool boundary

Diagnostic pass · not fixed

Hypothesis

The tool works, but the original call used the wrong argument contract.

Tool arguments

Original

{ query: "France" }

Controlled branch

{ country_code: "FR" }

Tool result

Recorded

[]

Live tool

["Paris"]

Execution provenance

Prompt: Recorded
Tool arguments: Edited
Tool result: Live
Final response: Live
Outcome contract: Passed deterministic assertion

Original arguments: 0 / 3 passed
Corrected arguments: 3 / 3 passed

Bounded finding

The live tool consistently returns valid data with country_code. The fault boundary has moved to the contract or the code, prompt, or state that formed the original call.

Next focus: Change the production tool contract, then rerun the original scenario without overrides.

Fix v1

Verify the production change

Verified

Production change

A required country_code contract should make the agent form the correct call without intervention.

Tool schema

Before

query?: string

After

country_code: string

Agent call

Failed run

{ query: "France" }

Real rerun

{ country_code: "FR" }

Execution provenance

Prompt: Live
Tool arguments: Live
Tool result: Live
Final response: Live
Outcome contract: Passed no overrides

No-override runs: 3 / 3 passed
Exploratory overrides: None

Bounded finding

The schema change makes the original scenario pass through the real agent with no exploratory overrides.

Next focus: Keep the verified scenario as a rerunnable regression tied to Failure #42.

Regression · no overrides

destination_suggestions_not_empty

Passed

Source: Verified no-override run. The production change—not either diagnostic branch—is now verified and protected.

Keep observability. Add a debugger for the hard failure.

Keep the tools that detect and monitor issues. Use Debrix when one failure needs a controlled local investigation.

Existing stack

Monitor production behavior, detect patterns, cluster failures, and measure impact. Keep the systems that already do this well.

Debrix

Investigate failures locally. Freeze evidence, compare controlled branches, and verify the real fix separately.

From failure to verified fix

One workspace carries the original evidence, every diagnostic branch, the production change, and the regression that keeps it fixed.

Failure
Own the failed execution
Preserve the original evidence, the violated expectation, and every investigation decision in one durable Failure.
Reproduce
Hold the starting point still
Reconstruct the baseline from recorded evidence so nondeterminism cannot move the failure before you test a hypothesis.
Modify
Run a controlled branch
Change one event or input, label every recorded, edited, and live step, and use the outcome to narrow the fault boundary.
Prevent
Verify the actual fix
Rerun the real code, prompt, schema, or policy change without exploratory overrides, then save the passing case as a regression.

Built for your runtime

Evidence you can inspect. Success you define.

Debrix owns the experiment plan and provenance. Your agent, codebase, and approved validator stay in control.

Local, in your runtime: Evidence and experiments stay local while your existing agent and codebase continue to run normally.
IDE-native: Cursor, Claude Code, and Codex can inspect evidence and run your project’s existing workflow over MCP.
Evidence-first: Every branch keeps provenance. Your test, assertion, human verdict, or labelled judge defines success.

Ready for a handful of agent engineers.

Debrix is ready for controlled early access. We’re onboarding agent engineers in small batches so we can stay close to the debugging workflows that matter.

There’s no public desktop download yet. Join the waitlist and we’ll reach out as we open the next onboarding batch.

Selected engineers get their first three months of Pro free.

Instrument once. Debug from your IDE.

The desktop is gated. The instrumentation and IDE workflow are open, so your coding agent can inspect evidence and drive your existing run.

Python SDK
pip install debrix
github.com/goyal-prateek/debrix-sdk
IDE skills & MCP
Cursor, Claude Code, and Codex skills for instrumentation and debugging via MCP.
github.com/goyal-prateek/debrix-plugin

FAQ

Straight answers about the product and controlled early access.

Debrix is a local-first interactive debugger for AI agents. It turns one failed run into an investigation with frozen evidence, controlled branches, a verified production fix, and a regression test.

Bring the next failure to Debrix.

Join the waitlist for controlled early access to the complete local Failure → Fix loop.

The interactive debugger for AI agents

One failure. Multiple controlled branches. One verified fix.

Test the tool boundary

Keep observability. Add a debugger for the hard failure.

From failure to verified fix

Own the failed execution

Hold the starting point still

Run a controlled branch

Verify the actual fix

Evidence you can inspect. Success you define.

Ready for a handful of agent engineers.

Instrument once. Debug from your IDE.

FAQ

What is Debrix?

Does Debrix replace my observability platform?

How does controlled debugging work?

Does a passing diagnostic branch mean the failure is fixed?

How does Debrix know whether a run succeeded?

Does Debrix require a custom agent runtime?

Can I download Debrix today?

Bring the next failure to Debrix.