Writing
Notes on building Envelope, the AI agent generator — architecture decisions, market observations, and ideas we're working through.
AI agents for sales teams: what to build and where to start
Sales has the highest density of repetitive, structured workflow work — prospecting, enrichment, outreach, CRM hygiene, and pipeline reporting. Five worth building as AI agents, and what to get right before you start.
ReadAI agents for support teams: what to build and where to start
Support has the highest ticket volume of any customer-facing function, and most of it is structured, repeatable triage work. Five workflows worth building as AI agents, and what to get right before you start.
ReadAI agents for marketing teams: what to build and where to start
Marketing has the highest volume of structured, repeatable content workflows — research, drafting, scheduling, and reporting. Five worth building as AI agents, and what to get right before you start.
ReadAI agents for beginners: start here
An AI agent takes actions and completes tasks — it doesn't just answer questions. Here's what you need to understand to get started, without the jargon.
ReadDesign a multi-agent AI system without writing code
You don't need to write code to design AI agents. You need to know what you want them to do. Here's how to describe one in plain language — and what to hand off when you're done.
ReadAI agents for agencies: how to design and deliver them to clients
The agencies winning AI work right now aren't just running prompts. They're delivering designed, documented, hand-off-ready AI agents.
ReadSalesforce multi-agent automation: what to build and where to start
Salesforce holds your pipeline. Multi-agent AI teams do the work to keep it moving — researching accounts, scoring leads, monitoring stalled deals, and flagging renewal risk.
ReadAI agents for Zendesk: how to build a smarter support operation
A single AI bot connected to Zendesk can categorise tickets. AI agents can triage, route, draft, escalate, and review — without anyone having to manage the queue.
ReadHubSpot AI agents: what to build and where to start
Most HubSpot AI automation stops at one-shot tasks. AI agents split the work between specialist agents — a Researcher, a Scorer, a Router — each with one job and a clean handoff.
ReadThe anatomy of a multi-agent system
What a multi-agent actually is, the parts it's built from — sub-agents, model routing, pipelines, gates — and why most attempts to build one collapse without a design step.
ReadThe AI agent integration problem
One working agent. A second bolted on. A third. Then everything breaks at the joins — overlapping instructions, undefined handoffs, no gates. Why multi-agent systems fail at the connections, not inside them.
ReadWhat is AI workflow automation?
Traditional automation connects apps and runs sequences. AI workflow automation handles the steps that require judgment — reading, deciding, routing based on context. Here's how the two differ, and what good multi-agent workflow design actually looks like.
ReadModel routing in multi-agent workflows
Different agents in the same system should run on different models. A classifier doesn't need the same model as a drafting agent. Here's how to match each role to the right model — and why it matters for cost and quality.
ReadHow to design an AI workflow before you build it
Most multi-agent systems fail because they were assembled, not designed. Here's a step-by-step process for going from a workflow objective to a complete spec — agents, handoffs, gates, and model assignments — before writing a line of code.
ReadHuman-in-the-loop: how to design approval gates into a multi-agent workflow
A gate designed in from the start is a checkpoint. A gate added after the fact is a workaround. Where to place gates, how to specify approve/reject/timeout behaviour, and a worked example before any code is written.
ReadSub-agent design: the five components every agent needs
Role, skills, model, tools, and handoffs — what each component covers, how they interact, and how to write a complete agent spec before any code runs. With a worked example.
ReadEnvelope vs. n8n, Make, and Zapier: what's the difference?
Envelope is the design layer for multi-agent AI systems. n8n, Make, and Zapier are execution layers. They solve different problems — and in most serious AI workflows, you'd use both.
ReadParallel vs. sequential agents: when to use each
Run agents in parallel when their work is independent. Run them in sequence when one genuinely needs the other's output. The pipeline structure follows from the data dependencies — not from a preference for speed or simplicity.
ReadMulti-agent AI: the complete design guide
Everything you need to design a multi-agent AI system — what multi-agent means, when to use it, the five design decisions that determine whether it works, and the mistakes that cause most teams to fail.
ReadHow to write a good AI agent role definition
A prompt tells an agent what to do. A role definition tells the system where one agent ends and another begins. How to write agent roles that produce reliable, predictable behaviour — and why the distinction matters more when a team is published for others to install.
ReadHow to design AI agents: a practical guide
Designing AI agents is a distinct skill from prompting a single model. Here are the six design decisions that determine whether an AI agent system works — and how to make each one before you build.
ReadAI agents for finance teams: what to build and where to start
Finance workflows are structured, high-stakes, and repetitive — the ideal conditions for AI agents. Five workflows worth building as agents, and how to design them with human oversight built in.
ReadHow to test an AI agent before deploying it
Testing an AI agent isn't like testing a function. Outputs are non-deterministic, tool calls have side effects, and failures can cascade silently. Here's how to test each layer before anything goes to production.
ReadWhen to use one AI agent vs. many: a practical decision guide
The instinct when designing AI agent systems is to add more agents. Usually that instinct is wrong. Here's how to decide when a single agent is enough and when splitting actually helps.
ReadAI agents for HR teams: what to build and where to start
HR is full of structured, repeatable work that slows down when done manually at scale. Five workflows worth building as AI agents, with human oversight designed in from the start.
ReadAI agent security: what to lock down before you deploy
AI agents act autonomously, chain together, and reach external services. Here's what to lock down at design time — before a misconfigured agent causes real damage.
ReadWhat are AI agents? A plain-language explanation
An AI agent is software that can take actions, use tools, and make decisions to complete multi-step tasks — not just answer questions. Here's how they work and what makes them different from chatbots.
ReadAI agents vs chatbots: what's the actual difference?
Chatbots respond to a single input. AI agents execute multi-step tasks using tools, memory, and autonomous decision-making. Here's how to tell them apart — and when to use each.
ReadAI agent handoffs: how to design clean transitions between agents
A handoff is the contract between agents — what one produces and what the next expects. Most multi-agent systems fail here. Here's how to design handoffs that don't break.
ReadAI agents vs automation: when to use each
Automation follows rules. AI agents make decisions. They solve different problems — and knowing which to reach for, and when to combine them, is the key design question.
ReadAI agent memory: how agents remember context across tasks
Agents don't remember by default — memory is something you design in. Here are the three types, when to use each, and the design decisions that separate a useful agent from one that keeps starting from scratch.
ReadAI agent observability: how to monitor agents in production
Testing tells you an agent works before you deploy it. Observability tells you whether it's still working after. Here's what to trace, measure, and alert on in production.
ReadAI agent spec to working software: how to close the gap
You designed the agent system. Roles, tools, handoffs, reporting lines. Here's how to close the gap from a clean spec to running code.
ReadAI agent costs: how to estimate and control your spend
AI agent costs have three components: model inference, tool calls, and human time at gates. Here's how to estimate each and the levers that bring spend down.
ReadAI agents need managers too
When an AI agent does something unexpected, the first question isn't technical — it's who is accountable. The answer depends on whether you built the management layer, or skipped it.
ReadThe model path won't solve coordination
Everyone is chasing better models on the assumption that coordination will follow. It won't. The execution layer isn't a workaround for model limitations — it's the governance layer.
ReadThe composable AI thesis
Any agent role can be decomposed into a team of narrower agents without changing the interface upstream. The agent abstraction is fractal.
ReadNarrow by design: the case for composable AI agents
The agents that perform best are not the most capable ones. They are the most focused ones. Why narrow roles outperform broad ones.
ReadHuman in the loop and the autonomous agent problem
The demand for human oversight is real. The infrastructure to support it properly is catching up. Here is why approval gates are harder than they look.
ReadCredentials, trust and the AI operator problem
Three parties, one secret. When builder, deployer, and runtime are different actors, who holds credentials — and under what conditions?
ReadThe platform trap in AI
When you build AI agents on a platform today, you're not just choosing infrastructure — you're choosing a format. The format is the lock-in.
ReadWhy open standards matter for AI infrastructure
The internet runs on standards nobody owns. AI infrastructure doesn't — yet. What format lock-in actually costs, and why the ecosystem effects of an open standard can't emerge from a closed one.
Read