Recent content on javatask.dev
Once app, config, weights, and tools are digest-addressed artifacts with attestations attached, much of the evidence the 2026-2027 EU obligation stack asks for becomes a query against infrastructure that already exists — not a compliance pipeline someone builds. The registry is the evidence trail; it is never the judgment that the trail was adequate.
An agent's behaviour is set by its container, its weights, and its tools. Model signing now has a credible, versioned answer — Sigstore model-transparency v1.0. Tool-definition provenance does not: as of July 2026, the official MCP Registry verifies who published a server, not what is in it.
OCI registries already let a second artifact — a signature, an SBOM, an attestation — attach to a first one without touching its digest. As of July 2026, most registries only implement half of that mechanism, and almost no edge deployment uses the half that exists.
One OCI runtime ships containers, Iceberg data tables, and model weights to a single on-prem box — the closed-system edge stack for a sovereign, AI-native factory with zero cloud round-trip.
Field notes from Margo Plugfest #2: what a multi-vendor round-robin proved about the Helm Way OCI packaging pattern, why the failures were the deliverable, and the architectural case for a packaging format nobody owns.
Each edge vendor's 'just one container' is a rational local optimum. The shop floor sums them into ten management planes. Margo lets every container land on shared edge infrastructure.
Three concrete architectural deliverables the EU regulation stack demands: Article 12 bitemporal traceability logs, NIS2 24/72/1-hour incident-reporting pipeline, and CRA SBOM generation in CI/CD. Where Apache Iceberg and OPA/Rego fit the implementation.
Every layer of the interoperable, agentic industrial edge now has an open, neutral-governance standard behind it. The full stack, layer by layer, with the governing body for each.
Gartner forecasts 80% of custom edge software containerized by 2028. The question is not whether to containerize — it's which runtime survives a brownfield OT floor, an air-gapped network, and a team with no Kubernetes operators.
Interoperability is a standards choice. Margo's pull-based desired-state model (port 443, brokerless) bridges legacy fieldbuses and breaks vendor lock-in without a central orchestrator.
Golden-image pull plus OCI ModelKits put apps and models on one signed, versioned, auditable lifecycle. MTTR collapses and agents become reproducible.
On-box GGUF/NPU inference removes the WAN dependency and satisfies OT data-residency requirements. Host-based offline licensing keeps the air gap intact under IEC 62443. Here is the mechanism and the pattern.
~88% of enterprise agent pilots never reach production (IDC). The problem is not the model — it is the missing deterministic harness. This post builds three concrete harness primitives: a read-only governance gate, a Plan-Execute-Verify loop, and a constrained context graph, then shows how they run on the sovereign Quadlet substrate.
OT telemetry is structurally chaotic until you layer two things on top of it: a Unified Namespace via MQTT for real-time context, and an Apache Iceberg lakehouse for queryable history. Together they form the shared substrate that analytics and agent memory both depend on.
Most writing about LLMs in industry has no number attached. This one does: one everyday operation — refilling a parts bin — instrumented end to end and costed to the cent. A vendor-published sensor-and-SaaS kanban system costs approximately $544 per bin over three years; the AI-native cell on my desk lands at approximately $40.50. Both figures are real, and the difference between them is…
One $4,000 industrial edge server manages a 20-cell factory and 400 bins with no container-orchestration layer on the box and no per-bin power. The 1:20 camera-to-bin ratio is the economic lever that makes autonomous e-Kanban viable.
Counting metallic fasteners under factory lighting is intractable. A Color-Blocked Ramp floor liner and AprilTag SKU marker solve fill-level detection for under $0.50 per bin — no powered sensor, no electronics, no wiring.
The duty cycle fires 720 times per day per cell. Where the reasoning model runs — and whether it runs on every cycle or only on genuine threshold crossings — is the single architectural decision that makes or breaks the TCO.
The vision-language model is open-weight. The local LLM is commodity. What makes autonomous e-Kanban trustworthy on a factory floor is the deterministic harness: the BeforeToolCallEvent provenance gate, the bounded belief update, and the Episode lineage queryable via Apache Iceberg and AWS S3 Tables.
The three-year TCO for 400-bin autonomous e-Kanban: approximately $40.50 per bin over three years, measured against the honest market baseline — vendor-published sensor-and-SaaS pricing that brackets $168–$544 per bin depending on tier. The spread between the two cost structures is roughly 4×–13×, with $0 recurring cloud cost. Here is the cost decomposition, the real vendor anchors, and what…
DevConf.CZ 2026's programme read as one converging argument: AI produces artifacts, signing proves them trustworthy, image-mode Linux ships them immutably, OpenTelemetry observes them. The next step is a loop that feeds itself.
The three protocol-level attack vectors A2A introduces — session smuggling, Agent Card poisoning, and SSE stream hijacking — and private-CA-signed TLS inside a sovereign VPC as the structural air-gapped mitigation. As of mid-2026, A2A spec v1.0.
Bi-temporal edges — valid-time and transaction-time on every relationship — turn an agent's memory graph from a faster lookup into an auditable record you can query as-of any past moment, without semantic approximation.
On Bedrock, Digital Coworker's read-only guarantee was the absence of a write Lambda in the Action Group registry. Porting to Strands for the air-gapped target, the lifecycle-hook system rebuilds that deterministic read-only seam at the SDK level — as of mid-2026.
On the cloud a managed scheduler gives you durable execution for free; air-gapped, you rebuild persist-replay-resume, liveness monitoring, and idempotency gates on a self-hosted state store — the substrate that replaces the managed orchestrator.
The runtime control loop that turns an agent that loops forever into one that finishes — constrained flow control, the Plan-Execute-Verify sandwich, and the Ralph context-recovery loop.
Harness policy externalised as inspectable, version-controlled plain-language documents a runtime interprets into deterministic actions — an emerging pattern I've been running, anchored in AGENTS.md and CLAUDE.md practice.
The payoff of getting packaging right: a new runtime drops in for almost nothing. Quadlet runs a container as a native systemd service — no Kubernetes, no daemon — on the constrained OT devices that can afford neither. Under an open Margo proposal (SUP-02, not yet adopted spec), adding it leaves the app developer and registry untouched, costs the fleet manager one schema-additive patch, and only…
The proposed fix for all three Compose failures is one move Helm already made: stop hosting a file, start pushing an artifact. A Margo TWG proposal (SUP-01, open for review) would publish the Compose Archive to an OCI registry with oras push, reference it by repository + revision, and let integrity fall out of the digest for free. Here is the before-and-after — and why docker compose publish…
Every way to publish a component must answer five questions: Storage, Reference, Media Types, Integrity, Parametrization. Helm answered all five years ago — through the OCI protocol, not a spec clause. Naming them as a checkable conformance contract is what lets Compose and Quadlet inherit the pattern instead of reinventing it.
How a community of 40+ organizations went from 'not a single example of Margo' to a live multi-vendor demo at SPS 2025 — and what the quadlet proposal is still trying to resolve.
A Margo Compose app today is a tarball at a URL, its signing key typically at another URL on the same server. Nothing lets the receiving device prove the file it pulled is the file the author published. Here are the three ways that bites — none exotic, all reproducible against today's Pre-Draft.
A single Margo application can ship the same service two ways — as a Helm chart and as a Compose Archive — and the spec publishes those two ways through two different worlds. One is content-addressed and OCI-native; the other is a tarball at a plain URL whose signing key is fetched over that same unauthenticated channel. That asymmetry is the last-mile gap this series closes.
The preconditions Phase 3 constrained write requires in Digital Coworker — deterministic validation layer, diff-level audit, blast-radius constraints, and approval boundaries.
The Margo specification has shipped the canonical OCI component publishing workflow from day one — in Helm. Naming that pattern is what lets Compose and Quadlet compose onto the same substrate instead of reinventing it.
The Computer&Automation Produkte des Jahres 2026 readers' choice drew 19,300+ votes across 96 nominated products, putting industrial readers in the role of technical evaluators. The questions embedded in our category placement and the coverage it generated were more specific than any internal review we had run.
The canonical post-talk write-up of my DevConf.CZ 2026 Brno talk: why Margo + Quadlet is a spec-driven stack for statically configured production cells, how it sits between apt/dnf+Ansible, Docker Compose, and Kubernetes, how the same `.container` descriptor targets WASM tomorrow, and what runs on top in production.
AWS Bedrock Agents externalises agent instructions into a console interface non-engineers can reach. What that means in practice for an OT diagnostic product — and what it does not solve.
When three AI agents reason over the same Apache Iceberg tables, the architecture — not each agent independently — guarantees consistency. This is what epistemic ground means in a multi-agent industrial system, and why the data layer decision is actually an agent-trustworthiness decision.
How one anti-hallucination principle — the LLM owns interpretation, deterministic code owns calculation — shapes every layer of the Digital Coworker architecture from data ingestion to safety enforcement.
OT network management has a structural knowledge-scarcity problem. This post explains the architecture designed to solve it — and what 19,300 industrial practitioners said about it.
In Digital Coworker, read-only agent posture is not a policy setting — it is enforced by the absence of write Lambda functions in the Bedrock Action Group registry. No tool exists to call, so no prompt injection can trigger a configuration change. Here is why that matters in OT and how the trust-earning sequence works.
A single AI agent with both VLAN and MRP instruction context degrades at both. The Digital Coworker architecture uses three specialised agents — VLAN, MRP, Network Analysis — and routes between them. Here is the mechanism.
Four observable architectural decisions define AI-native vs AI-augmented. One month after publishing that claim, here is which markers proved durable under production pressure and which need qualification.
RAG is the default answer for grounding LLMs in domain data. For OT network topology, it was the wrong choice. An empirical comparison on a 10-device mesh and the architectural principle that generalises.
Fan-out / fan-in dispatch is an economics problem before it is an architecture problem. Part 2 of the Agentic AI on AWS series — what changes when you compose multiple Kiro subagents, why isolated context is an architectural property not a workaround, and what cost-optimization MCPs are actually for.
What changes when requirements.md, design.md, and tasks.md become the contract your AI agent delivers against — instead of a chat history that evaporates between sessions. Part 1 of the Agentic AI on AWS series, anchored on AWS Kiro CLI.
An AI diagnostic agent placed second in a category dominated by switches and embedded modules. That it competed directly — not in a separate AI category — tells you where the industrial automation market is in its transition.
Most AI-native claims are retroactive relabeling. Three observable architectural differences that distinguish genuinely AI-native products — drawn from the Digital Coworker multi-agent OT diagnostic system.
Margo v1.0 specifies Helm and Compose deployment profiles. On the sub-512 MB device tier — below the k3s server minimum, where most OT gateways sit — neither fits. Podman Quadlet is already there, on every RHEL, Fedora, and Ubuntu edge image.