Blog
Notes on coordinated intelligence — how autonomous systems earn operational trust.
Machines work better in tandem
Models can reason. Tools can act. The missing layer is the system that coordinates both — while keeping people informed, accountable and in control. Our founding argument, in full.
Read the thesis →Agent washing: how to tell a real agent from an AI wrapper
Half the products sold as AI agents are a chatbot with a system prompt. Here is a test battery you can run in a demo — five questions whose answers separate a real agent from a wrapper, no matter what the marketing says.
AI agent design patterns: ReAct, plan-and-execute, and reflection
The three core agent patterns are not interchangeable — each makes a different bet about when the model should think, act and check. Here is what each is for, and the failure each one is quietly prone to.
Agentic RAG: retrieval with a decision loop
Classic RAG retrieves once and answers. Agentic RAG lets the model decide what to retrieve, judge whether it got it, and try again — which turns retrieval failures from a dead end into something that compounds.
Context engineering: the discipline that replaced prompt engineering
Prompt engineering is writing good instructions. Context engineering is deciding what occupies a finite window and what it costs — and once you understand the KV cache, prefix stability stops being a style choice and becomes an economic one.
A prompt injection catalogue: working examples and what stops each
One entry per attack: the payload, the channel it arrives through, what it achieves against an agent that can act, and the control that actually stops it — which is never a filter.
Race conditions in agent loops: parallel tool calls and the state you didn't lock
Agents emit parallel tool calls by default, and a language model has no concept of a transaction boundary. Two calls in one turn can write the same row — and the pattern you chose determines which races you get.
Calibrating an LLM judge: inter-rater reliability for model-graded evals
An uncalibrated LLM judge is an opinion with a decimal point attached. Measure agreement with Cohen's κ before you trust it — and measure your humans against each other first, because that is the ceiling.
What are agentic workflows — and when a workflow beats an agent
If you can enumerate the steps in advance, you have a workflow — and a workflow is cheaper, testable, auditable and incapable of surprising you. Most systems sold as agents are workflows, and most teams should build one.
After-hours answering: the unit economics of never missing a call
After-hours answering is not a cost-reduction case, it is a revenue-recovery case — and that changes the arithmetic completely. Here is the model, the break-even containment rate, and the cases where the honest answer is that it does not pay.
Agent authority as code: writing the boundary in Rego
An agent is not a principal — it is a delegate. That single fact makes role-based access control the wrong model, and it is why authority has to be expressed as policy evaluated on the tool call rather than as a sentence in a prompt.
Hermes Agent vs OpenClaw
This is not a symmetric comparison. Hermes ships a first-party `hermes claw migrate` command that imports OpenClaw's memories, skills and config — which tells you more about how these two relate than any feature table.
The voice agent stack: latency, turn-taking and barge-in
Your voice agent's response time is dominated by how long it waits to decide you stopped talking, not by inference. That is a configuration decision, and it is the one that makes a call feel human or not.
Mastra vs LangGraph vs CrewAI
The three differ on one axis: what they force you to make explicit. A team, a state machine, or a TypeScript application. Plus the LangGraph resume behaviour that will duplicate your side effects if you haven't read the tests.
Semantic caching for AI agents
Semantic caching is designed for repeated questions, and an agent loop almost never repeats a prompt. The wins are on the tool side — and the cache key has to carry the requester's authorization scope, or the cache becomes a data-leak channel.
Choosing the agent data layer: pgvector vs Chroma vs Pinecone
Start in Postgres. Here are the specific thresholds where that stops being true, why an agent's retrieval pattern argues for co-location, and why your recall problem is almost certainly chunking rather than the database.
Agent observability: Langfuse vs Arize Phoenix vs Braintrust
Before choosing a tool, decide what a trace has to contain to answer “why did it do that?”. Then one architectural question — OpenTelemetry-native or proprietary SDK — narrows the field faster than any feature matrix.
The AI agent tech stack: seven layers
Most stack diagrams stop at the control loop. The two layers after it — governance and evidence — are what separate an agent you demo from an agent you deploy, and they are architecture, not process.
Idempotency for AI agent actions
A retried HTTP request is a duplicate. A retried agent action is a second refund. Every idempotency pattern you know assumes a client that repeats itself exactly — an agent does not, and that breaks the key.
Building production agents with the Claude Agent SDK
The SDK is Claude Code as a library. This is the part the quickstart skips: the six-step permission evaluation order, the three ways your approval callback gets silently bypassed, and what a subagent actually inherits.
Agentic AI security: a working threat model
An agent with tools turns a prompt injection into a transaction. The threat model for systems that can act, the three conditions that make exfiltration possible, and the controls that hold when the model is fooled — because it will be.
Agentic AI vs generative AI vs AI agents
Three terms, used interchangeably by people who should know better. Here is what each one actually denotes, why the major vendor definitions contradict each other, and the two questions that tell you which you are dealing with.
What is a forward-deployed engineer?
The role Palantir invented and every serious AI-agent company is now hiring for — what an FDE actually does, how the job differs from solutions engineering and professional services, and why agent products cannot be delivered without one.
When should an agent ask?
Designing the authority boundary between machine execution and human judgment.
Evidence is the interface
Why every agent action should leave a record a person can review — and what that changes about trust.
One objective, four stages
Direct, coordinate, execute, verify — the anatomy of a Tandem run, from instruction to verified outcome.
The case against parallel tools
A drawer full of disconnected copilots is not an operating system. Why coordination is the unit of value.
Follow the work.
One essay a month on coordination, control and operational AI. Nothing else.
By subscribing, you agree to the Tandem Privacy Statement.