Writing
Blog
Technical articles, benchmarks, and insights about building reliable AI systems with OpenSymbolic.
- April 13, 20264 min read
DesignExecute: When Straight-Line Plans Aren't Enough
PlanExecute forbids loops and conditionals on purpose. DesignExecute adds them back, with guardrails, for the problems that actually need control flow. Here's when to reach for it, and what stays the same.
architectureDesignExecutePlanExecuteagents - March 24, 20264 min read
Third Language, Same Result: MultiHopRAG in Go
Go joins Python and C# on the MultiHopRAG benchmark. Different runtime, different vector store, single static binary. Accuracy: 81.6%. The framework holds.
benchmarkcross-languagecode-generationagents - March 23, 20266 min read
Change Everything, Change Nothing: MultiHopRAG in Python and C#
We swapped the language, the vector store, the code executor, and the type system. Accuracy moved by 0.9pp. The framework is the invariant, not the infrastructure.
benchmarkcross-languagecode-generationagents - February 20, 20265 min read
Agent-to-Agent Is Just Function Calls
Multi-agent systems don't need new infrastructure. They use the same patterns that connect microservices today: typed interfaces, explicit wiring, and the auth and observability stack you already have.
architectureagentsa2amulti-agent - February 20, 202612 min read
TravelPlanner Benchmark: 97.9% on 1,000 Tasks Where GPT-4 Gets 0.6%
OpenSymbolicAI achieves near-perfect scores on all 1,225 TravelPlanner tasks, outperforms LangChain and CrewAI head-to-head, and maps the model landscape across 11 LLMs and 4 providers.
benchmarkcomparisonagentsreliability - February 19, 202615 min read
The Anatomy of PlanExecute: Why It Is What It Is
A deep dive into the design decisions behind OpenSymbolicAI's core blueprint: why plans are code, why execution is sandboxed, and why the planner only gets called once.
architecturePlanExecuteagentsdesign - February 19, 202613 min read
English, Spec, or Code: How You Talk to the LLM Decides How Far You Get
Plain English gets you to 50% on day one. Specs get you to 70%. Code gets you to 99.999%. The question is which ceiling you're willing to live under.
opinionarchitectureagentsprompt-engineering - February 3, 202611 min read
Secure by Design: How OpenSymbolicAI Makes AI Agents Safe
Why security in OpenSymbolicAI is architecturally guaranteed, not hoped for. Introducing the Symbolic Firewall.
securityarchitectureenterprisecompliance - February 1, 20264 min read
Illustration Part 3: Cost & Reliability
Tool-calling costs 2.3x more and fails 20% of the time. Here's the math.
agentsillustrationperformanceRAG - February 1, 20263 min read
Illustration Part 2: Token Economics
Tool-calling re-reads everything on every call. Here's exactly where the tokens go.
agentsillustrationperformanceRAG - February 1, 20264 min read
Illustration Part 1: The Attention Loss Problem
We gave a tool-calling agent detailed instructions. It ignored them. Here's why.
agentsillustrationperformanceRAG - February 1, 20266 min read
LLM Attention Is Precious: Why ReAct Wastes It
A visual comparison of token usage between ReAct (Reasoning+Acting) and PlanExecute approaches.
agentsarchitectureperformancetool-calling - January 30, 202613 min read
Behaviour Programming vs. Tool Calling: Two Paradigms for AI Agents
Why teaching agents through executable examples beats massive prompts with chain-of-thought reasoning, illustrated through a RAG agent implementation.
agentsarchitectureRAGbehaviour-programming - January 28, 20266 min read
Closing the Flywheel in Practice
A hands-on walkthrough showing how primitives and decompositions create the compounding improvement loop that prompt engineering lacks.
tutorialagentsprimitivesdecompositions - January 28, 20268 min read
The Missing Flywheel in Agent Building
Why language models keep getting better while AI agents remain stubbornly brittle, and how to fix it with structural separation of concerns.
agentsarchitecturereliabilityphilosophy - January 27, 20264 min read
The OpenSymbolicAI Manifesto
How do we build non-deterministic intelligence on top of a deterministic foundation, without compromising either?
manifestophilosophyarchitecture