← Back to projects

ANONYMIZED CASE STUDY / PRIVATE SOURCE

Reviewable Runtime

An evidence-driven agent runtime built to be inspected, interrupted, and resumed

ROLE

Python Agent services and the Agent core: APIs, context assembly, controlled execution, capability governance, evidence ledger, checkpoints and replay, temporal memory, offline evaluation, and Java/Python contracts.

SCOPE

The public case excludes the original name, organization, screenshots, source code, and client context. It does not claim ownership of the frontend, Java business implementation, or the whole product.

Long-running agents combine model decisions, tools, user data, and asynchronous work. Without explicit runtime boundaries, a plausible answer can lose its source, an interrupted task can become unrecoverable, and a model request can quietly turn into an unauthorized action.

Runtime boundary mapThe highlighted runtime is the part I worked on.
01Web client
02API gateway / BFFidentity · files · session
03 / OWNEDPython Agent runtimecontext · policy · tools · evidence
04Durable executionqueue · checkpoint · event log
↳ Authoritative store
↳ Temporal memory projection
01

Evidence before claims

The flow is evidence → finding → claim → report. A report can only use findings backed by recorded evidence; missing evidence degrades the result or requests human input.

02

The model proposes; the runtime decides

A deterministic capability gateway owns permissions, feature flags, budgets, timeouts, retries, and approval. The model never becomes the authority just because it produced a tool call.

03

Recovery is a product feature

Background jobs, checkpoints, and a durable event log let work pause, resume, and replay after disconnects. Streaming is treated as a view over durable state, not the state itself.

04

Memory remains attributable

The relational store is authoritative. Temporal memory is a rebuildable projection with provenance, validity, and withdrawal semantics rather than an opaque bag of model-generated facts.

The snapshot also measured a 100% Hybrid Recall@5 on 30 synthetic interview-knowledge queries, while inference leakage, withdrawn-information leakage, and unauthorized memory writes remained at 0%. These are controlled offline results, not production usage claims.

95.24%Context Recall@1036 gate cases
100%Context Precision@1036 gate cases
100%Provenance completenessoffline snapshot
0%Unauthorized memory writesoffline snapshot

!The full evaluation was not all-green: one question-extraction invoke golden drifted. The result is shown because failure boundaries are part of the evidence, not something to hide.

This case intentionally omits the original repository, product name, organization, screenshots, package names, and infrastructure identifiers. The architecture is redrawn at the responsibility boundary. Metrics come from an offline snapshot dated 2026-09-15.

A useful agent should be able to answer a second question:

“Why should I trust this result?”