October 24, 2026ENGINE
Declarative JSON Workflow Engine — Initial Release
We shipped the core execution engine that parses declarative JSON DAGs and executes multi-step agentic workflows. Users define workflows as structured JSON with atomic steps (HUMAN, REASONING, EXECUTION), dependencies, and verification prompts. The engine handles state management, parallel execution, and error recovery.
Why it matters
Traditional agent frameworks require writing Python code to define workflows. Escepine uses declarative JSON so workflows are version-controlled, auditable, and safe for AI-generated logic.
The mechanism
The engine parses the JSON DAG, resolves dependencies, and executes each step with built-in verification. If a step fails, the engine triggers Saga compensations to roll back committed state.
October 23, 2026RESEARCH
18-Minute Unedited Stress Test — Resource Starvation Scenario
We published an 18-minute unedited terminal recording of the Escepine engine executing a 101-step workflow under intentional resource starvation. The test simulates throttled API limits, network degradation, and database volatility.
Why it matters
Most agent demos show the happy path. We show the sad path. The engine caught schema deviations, hallucinated tool calls, and stale data, then recovered autonomously.
The mechanism
The engine enforces hard iteration caps, validates schema at every tool call, and triggers Saga compensations when step execution fails. All events are logged to the immutable evidence ledger.
October 22, 2026POLICY
Cedar Policy Enforcement — Runtime Guardrails
We deployed the Policy Gateway, which evaluates Cedar policies at runtime before any tool execution. Policies are enforced outside the LLM context window, meaning the model cannot bypass, override, or reason past them.
Why it matters
Guardrails defined in prompts can be hallucinated past. Guardrails enforced at the runtime layer cannot. This is the difference between hoping the model follows rules and enforcing rules at the infrastructure level.
The mechanism
Before any EXECUTION step runs, the engine evaluates the associated Cedar policy. If the policy returns DENY, the step is halted and the workflow is parked for human review.
October 21, 2026RESEARCH
Lusser's Law and the Math of Agent Failure — Research Paper
We published our first research paper on the mathematical certainty of agent failure in long-running workflows. The paper demonstrates that even with 96% step accuracy, a 20-step workflow has a 44% probability of at least one failure.
Why it matters
This math is why execution infrastructure exists. Without Saga compensations and state management, every multi-step agent will eventually corrupt its state.
The mechanism
The paper introduces the Lusser's Law for Agents framework, which calculates the probability of workflow failure based on step count, tool reliability, and LLM hallucination rates.
October 20, 2026INFRASTRUCTURE
Technical Documentation and Security Architecture — Published
We published the complete technical documentation for Escepine, including the execution architecture, Cedar policy schema, Saga compensation patterns, and evidence ledger design. We also published the security architecture, detailing how MCP tools run in ephemeral sandboxes and how policies are enforced at runtime.
Why it matters
Enterprise buyers need to understand the architecture before they integrate. Documentation is the artifact the CTO forwards to the architect who will judge the integration.
The mechanism
The documentation is structured as a reference manual, not a marketing page. It includes architecture diagrams, API schemas, and failure mode analysis.
