DKTA Labs

For teams shipping autonomous software

Boundaries for
autonomous systems.

DKTA Labs builds the infrastructure around agents: execution, retrieval, evaluation, security boundaries, deployment, and operational feedback. The goal is simple—software you can inspect, constrain, and operate.

Explore the work Start a technical conversation

01 / What we build

The systems around the model.

Models generate. Reliable products still need ownership, state, permissions, evidence, and a release path. That surrounding layer is the work.

Execution
Local supervisors, durable run evidence, and explicit lifecycle ownership for scheduled agent work.
Retrieval
Machine-readable search, page extraction, OCR, and deterministic data utilities with bounded inputs.
Evaluation
Long-horizon environments that expose planning, tool use, adaptation, and completion behavior.
Release boundaries
Credential-blind dependency execution, reviewed artifacts, approval gates, and observable deployment proof.

02 / Selected work

Public artifacts, not pitch-deck claims

Working systems and inspectable contracts.

capsulectl

Public tool · pre-1.0

Credential-blind JavaScript dependency intake and execution. It builds reviewed dependency images inside disposable Docker boundaries and keeps promotion separate from dependency-controlled code.

Inspect the source

agentd

Maintained public source

A small local daemon for scheduled agent jobs, with SQLite-backed definitions and run evidence, one active run per job, and read-only operator visibility. It is maintained source, not a hosted service.

Inspect the source

TavernBench

Live arena · public clients

A long-horizon evaluation arena where agents navigate a real-time multiplayer world. The protocol, Python SDK, CLI, MCP integration, and Go TUI are public; the hosted server and held-out evaluation machinery remain separate.

Open the arena Inspect the clients

Extract

Live API · public source

Metered extraction and deterministic data utilities for agents: public-page extraction, bounded English-language image OCR, and JSON/CSV formatting. The service supports per-request x402 payments or a prepaid API key without a subscription.

Read the API surface Inspect the source

03 / Operating principles

Make consequential work legible.

The lab favors boring mechanisms that preserve ownership and evidence over invisible automation that merely looks fast.

  1. 01

    Durable state and explicit ownership over hidden process magic.

  2. 02

    Narrow trust boundaries and approval gates around consequential actions.

  3. 03

    Observable verification over claims that a workflow probably worked.

  4. 04

    Public client contracts separated from private operations and held-out evaluation machinery.

04 / Start with the actual system

What are you trying to make reliable?

Send the architecture, the constraints, or the failure you can reproduce. DKTA Labs is built and operated by Dakota Secula-Rosell.

dakota@dkta.dev