Skip to content
TP
← All work
Active build2025 – 2026

Tekkscope

Agentic search infrastructure that fails loudly instead of quietly.

PythonTypeScriptLangGraphVector searchRedisDockerCI harness
tekkscope.app

trace · task 88f2

4.2s · 6 calls
plan
search.web
search.index
ground
synthesize
stage ground below threshold · flagged in CI

success

0.86

tool prec.

0.91

cost/query

$0.03

Per-stage tracing

Plan, retrieval and synthesis are timed and scored separately rather than as one opaque run.

Failing stage

Grounding fell below threshold. The trace names the stage, so debugging is not guesswork.

Cost per query

Tracked as a first-class constraint rather than discovered on the monthly bill.

Hover a marker to see what each part of the interface does.

Task success rate

01

{{TASK_SUCCESS}}

Tool call precision

02

{{TOOL_PRECISION}}

Median steps per task

03

{{MEDIAN_STEPS}}

Cost per resolved query

04

{{COST_PER_QUERY}}

01

The problem

What was broken before any of this existed.

Agentic search demos look excellent and production agents drift. They retrieve confidently wrong sources, loop, or silently degrade after a model or index change.

Engineering teams putting agent-driven search in front of real users.

Constraints I worked inside

C01Non-deterministic components make traditional testing close to useless
C02Cost scales with every extra reasoning step
C03Failures must be attributable to a stage, not to "the agent"
90

ownership

end to end

50

role span

of four tracks

85

shipped

reached users

02

What changed

The same job, before and after the work.

Before

  • Agents drifted silently after model changes
  • One end-to-end score, no way to debug it
  • Unbounded loops discovered on the bill
  • Failures attributed vaguely to "the agent"

After

  • Retrieval, tool choice and synthesis scored separately
  • Fixed replay scenarios run on every change
  • Hard caps on planning depth and tool calls
  • Regressions surface in CI with the failing trace
03

The system, end to end

How a request travels through it.

PLANquery decomposed into 01ACTtool calls executed wi02GROUNDresults checked agains03SCOREreliability harness re04ALERTstage-level regression05
01

query decomposed into a bounded set of sub-goals

02

tool calls executed with per-stage tracing

03

results checked against source before synthesis

04

reliability harness replays fixed scenarios on every change

05

stage-level regressions surface in CI with the failing trace attached

04

Where the effort went

Effort split and how quality moved.

$stat --effort

Where the time actually went

effort
  • Reliability harness45%
  • Agent orchestration33%
  • Tracing and alerts22%

$plot --trend

How it moved

v1v2v3v4v5v6v7v8
task success rate across harness revisions, shape only
05

Decisions

Every call I made, and the tradeoff I took for it.

  1. Why

    Unbounded agents fail in the most expensive way possible: slowly, and with a large bill.

    Tradeoff accepted

    Genuinely hard queries get cut off rather than solved.

    Weighting

    gain
    cost

    my judgement, not a measurement

  2. Why

    A single end-to-end score tells you something broke but not what, which makes debugging guesswork.

    Tradeoff accepted

    More instrumentation to build and maintain before any of it pays off.

    Weighting

    gain
    cost

    my judgement, not a measurement

  3. Why

    Comparability matters more than realism when the question is whether this change made things worse.

    Tradeoff accepted

    The suite drifts from real usage and needs deliberate refreshing.

    Weighting

    gain
    cost

    my judgement, not a measurement

06

Outcome

What was true after it shipped.

01

Agent regressions became visible at the stage level in CI rather than after users noticed.

02

Cost per resolved query became a design constraint rather than a monthly surprise.

03

The reliability harness generalises to other agent products, which is where the work is heading next.

What worked

Treating agent reliability as an infrastructure problem instead of a prompt problem.

What was hard

Scenario design. Writing tasks that are hard enough to be informative but stable enough to compare took several rewrites.

What I would change

I would build the trace viewer earlier. Reading agent failures in raw logs slowed every debugging session.

$./next-step

Hiring for a product engineer who can carry the whole thing?

Send me the role and the problem behind it. I will tell you honestly whether I am the right fit.