Focus
Agentic workflows that survive real life
An agentic workflow is an engineered loop that interleaves model inference with tool use, state, memory, and feedback — letting a system accomplish work that exceeds a single prompt-response interaction. This page maps the stable patterns, the constraints that make them production-safe, and the artefacts I've shipped building them.
Workflows vs agents
Workflow
Predetermined control-flow written in code. You specify every branch, tool call, and retry policy explicitly. Predictable, debuggable, fast to test. The right choice when you understand the problem space fully and variability is low.
Agent
Dynamic control-flow where the model decides which tools to call, in what order, and when to stop. Handles variability and open-ended tasks. Requires bounded agency, observability, and human override mechanisms to be safe in production.
The practical answer: most production systems are hybrid. A fixed outer workflow handles routing, authentication, and budgets. Agents handle the variable inner work where predetermined branching would be brittle.
The core patterns
Routing
The model classifies an input and dispatches it to a specialised handler. Simple, predictable, fast — the right first pattern when you know your input categories.
Parallelisation
Fan out independent sub-tasks to multiple agents simultaneously, then aggregate results. Cuts wall-clock time when sub-tasks don't depend on each other.
Orchestrator–Worker
A planner agent breaks a goal into steps and delegates each to specialised workers. The orchestrator coordinates; workers execute tool calls. Best for open-ended tasks with variable depth.
Evaluator–Optimiser
One agent generates a candidate output; another evaluates it against criteria and feeds back. Loop continues until quality threshold or budget is met.
ReAct Loop
Interleave reasoning traces with tool calls: Reason → Act → Observe → Reason. The foundational loop behind most tool-using agents.
Human-in-the-Loop
Insert approval gates before high-stakes actions. The agent pauses, presents its intent, and only proceeds with explicit confirmation. Non-negotiable for production agents.
State, memory, and resumability
Long-running agents fail. Networks drop, processes crash, budgets expire. A workflow without checkpointing cannot recover — it must restart from scratch. Modern graph-based orchestration frameworks represent agent state as a serialisable graph that can be persisted after each step and restored on failure.
Checkpoint
Serialise the full agent state — tool results, decisions, memory — after every meaningful step.
Thread
A persistent, resumable conversation context keyed by run ID. Multiple humans or agents can interleave on the same thread.
Budget
Hard limits on steps, tokens, and wall-clock time. Without budgets, agents can loop indefinitely on edge cases.
Tool use and feedback loops
Tools are the agent's interface to the world: APIs, databases, filesystems, browsers. The loop is: model emits a tool call → application executes it → tool output returns to the model → model decides next step. Tool design is UI design for agents — poor descriptions and noisy outputs degrade reasoning quality.
Go deeper: AI Tooling hubMulti-agent collaboration
Roles separate concerns: a planner doesn't need to know how to write code; a code-writer doesn't need to know the full project plan. Handoffs transfer context between agents, with the receiving agent's role limiting what it can do. Agent-to-agent protocols (A2A) are now being standardised to enable interoperability across vendor boundaries.
Related: Sacred Technology hubStart here
Three things I've shipped that embody these patterns.
Agentic AI Landscape 2026: Toolsmith Field Guide
Infrastructure-first view of the agentic ecosystem — protocols, memory, orchestration, and what to build on.
What I Learned Building MCP Servers
Hard-won lessons on tool design, context poisoning, and making servers agents actually use.
Symbolic Ontology MCP (project)
An MCP server for symbolic reasoning — a concrete orchestrator-worker pattern in production code.
Common questions
What's the difference between a workflow and an agent?
A workflow uses predetermined control-flow: you specify every branch and tool call in code. An agent is dynamic — the model itself decides which tools to call and in what order, based on the current state and its goal. In practice, the best production systems mix both: a fixed outer workflow with agents handling dynamic sub-tasks where variability is expected.
When should I use an orchestrator–worker pattern?
When you have a goal that can be decomposed into steps, but you don't know those steps in advance. The orchestrator plans dynamically and can re-plan if a worker fails or produces unexpected output. Use it for research, code generation, or any task where the path to completion depends on intermediate results.
How do I make an agentic workflow resumable?
Checkpoint state after each meaningful step — tool call results, intermediate outputs, and agent decisions. Store checkpoints in a persistent store keyed by run ID (a thread). If the process crashes or times out, reload the checkpoint and continue from the last stable state. Treat every step as potentially the last.
What does human-in-the-loop actually mean in practice?
An approval gate that fires before irreversible or high-stakes actions: sending an email, deploying code, deleting data, or spending money. The agent presents its proposed action and waits. The human approves, rejects, or edits. Good HITL is surgical — you don't want approval gates on every tool call, only on actions where being wrong is expensive.
How do I test an agentic workflow?
Build an eval harness with 'golden' runs: a set of inputs where you know the correct sequence of tool calls and final output. Run your workflow against them on every code change. Flag regressions where the agent changed its tool selection or reasoning path unexpectedly. Evals are the unit tests of agent behaviour.
Also in Focus
AI Tooling →
MCP servers, tool calling, context engineering, observability, evals, and security guardrails.
Also in Focus
Sacred Technology →
Lineage-aware knowledge systems, calm technology, human–AI co-creation, and sacred persistence.
Building agents in production?
Let's talk architecture, tooling, and what doesn't work.