SELECTED ENGINEERING / PUBLIC EVIDENCE

Decisions behind the systems.

Problems, constraints, implementation choices, and the trade-offs that remain after a design becomes code.

These case studies examine public open-source work and research implementations. They do not represent confidential client projects or imply production adoption.

01 / Public open-source project

agent-guard

Making every tool call an authorization decision.

A policy boundary between what an agent proposes and what it is permitted to execute.

Agent security · Policy · Auditability

Problem & context

An agent can generate a valid tool request without having authority to perform it. A prompt instruction cannot establish that authority. The decision needs an explicit principal, action, resource, and context at the point of execution.

Constraints

Different runtimes need the same policy semantics. Delegated agents need bounded permissions, and operators need a traceable record of decisions without tying application logic to a single interface.

  1. 01Tool request
  2. 02Identity + context
  3. 03Cedar policy
  4. 04Allow / deny
  5. 05Audit record
A Rust core wraps the Cedar policy engine. SDK, CLI, console, and AuthZEN HTTP interfaces converge on policy evaluation and a hash-chained audit log. Versioned policy bundles separate policy changes from application releases.

Technical decisions

Keep authorization outside model reasoning. Share one evaluation core across integration surfaces. Represent delegation explicitly, with scoped and time-bounded tokens, instead of passing an unrestricted parent identity.

Trade-offs

A local engine avoids a network hop but distributes policy rollout across hosts. An HTTP decision point centralizes the boundary but adds availability and latency dependencies. Decision caching also creates a freshness window that must be part of revocation design.

Implementation

The repository separates core policy evaluation, authentication, telemetry, bundle management, CLI, and server crates. A TypeScript SDK and an administration console expose the same underlying decision path.

Reliability

The documented design includes policy validation, simulation, hot reload, and rollback. Treat denial, unavailable identity, and unavailable policy evaluation as explicit integration outcomes.

Security

Scoped delegation, token validation, sender constraints, and tamper-evident audit records address different parts of the trust boundary. The console requires OIDC configuration and fails closed when authentication is missing.

Evaluation

The public repository includes Rust and SDK tests, policy simulation, and an audit-chain verification command. An adopter still needs tests for their own principals, resources, revocation expectations, and failure modes.

Operating the system

Policy versions, authorization spans, cache lifetime, audit rotation, and signing-key management all require ownership. A tamper-evident log is useful only when retention and verification are operated deliberately.

Outcome & evidence

A public implementation of per-tool-call authorization with multiple integration surfaces. The repository demonstrates the mechanism; it does not establish customer adoption or a measured incident reduction.

What the design makes clear

The useful security boundary is where an action takes effect. Keeping that boundary explicit makes policy review, testing, and operational accountability easier to reason about.

Source: agent-guard public README. Architecture summary reviewed 21 September 2026. Operational considerations are analysis of the design, not reported deployment results.

02 / Public open-source project

promptsheon

Treating agent configuration as a release artifact.

Versioned configuration, evaluation, and approval belong in the same path to release.

Evaluation · Release engineering · Agent workflows

Problem & context

Changing a prompt or agent graph can change application behavior without changing application code. Teams need to identify exactly what ran, evaluate a candidate, control activation, and restore a known release.

Constraints

The platform is self-hosted and supports multiple model providers. Authoring, execution, evaluation, and approval need a shared identity for the configuration under review.

  1. 01Author graph
  2. 02Compile + hash
  3. 03Evaluate
  4. 04Approve
  5. 05Canary / rollback
A Fastify backend and Next.js interface use SQLite and a content-addressed store. Strands-based graph execution runs capability nodes; evaluation suites and maker-checker approvals govern releases.

Technical decisions

Store compiled manifests by content rather than mutable names. Separate release creation from approval. Make activation an explicit operation and retain the version needed for rollback.

Trade-offs

SQLite and local content storage keep the self-hosted deployment understandable, while making backup, concurrent access, and storage ownership explicit concerns. Automated scoring supports repeatability but cannot replace judgment about the right evaluation dataset.

Implementation

The platform provides a DAG editor, per-node configuration, provider integration, dataset-based scorers, environment activation, weighted canary releases, and an append-only audit chain.

Reliability

Release identifiers and persisted evaluation results support diagnosis and rollback. The documented self-evolution loop adds a separate control concern: automatically proposed changes still need bounded rollout and monitoring.

Security

Maker-checker rules prevent a release creator from approving their own release. Signed incoming webhooks and replay protection address the release integration boundary; deployment-specific access and secrets still need review.

Evaluation

Exact-match, regex, and model-judge scorers support different failure classes. Compare candidate and baseline on the same cases, and retain enough configuration to interpret a score later.

Operating the system

Back up the SQLite database and content store together. Track active releases, provider failures, evaluation coverage, and rollback behavior. A successful health endpoint does not prove output quality.

Outcome & evidence

A public platform that brings agent authoring, execution, evaluation, and release controls into one system. No claims about internal-team adoption or production improvement are made here.

What the design makes clear

Configuration is executable behavior. A release process becomes more useful when the evaluation result, approval, and deployed content refer to the same immutable artifact.

Source: promptsheon public README. Architecture summary reviewed 21 September 2026. Operational considerations are analysis of the design, not reported deployment results.

A GOOD CONVERSATION STARTS SOMEWHERE

Let’s connect.

Working on a difficult AI or engineering problem? Hiring for a senior architecture role? Or simply interested in exchanging ideas?

I’d like to hear from you