Practical, research-grounded writing on building agentic AI systems: architecture, diagrams, and runnable code.https://www.agenticsystems.ai/https://www.agenticsystems.ai/blog/a2a-explained/https://www.agenticsystems.ai/blog/a2a-explained/MCP connects an agent to its tools. A2A connects an agent to other agents: ones that think, take time, ask questions and belong to someone else. How A2A 1.0 works on the wire, and seven rules for using it well.Fri, 09 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/agent-identity/https://www.agenticsystems.ai/blog/agent-identity/Most agents run on a shared API key or a borrowed user login, so nobody can say who did what, and a confused agent can do anything. How to give agents their own identity, act for users through delegated tokens, scope access to the task, and keep a record that holds up.Tue, 03 Nov 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/agents-that-design-agents/https://www.agenticsystems.ai/blog/agents-that-design-agents/Since 2024, a line of research has had models write and improve agents, prompts and algorithms, with results ranging from better coding agents to a 56-year-old record in matrix multiplication. One loop explains all of it, and a 2026 paper explains why it's more fragile than it looks.Sun, 04 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/agents-that-use-computers/https://www.agenticsystems.ai/blog/agents-that-use-computers/Computer use lets an agent operate software the way a person does, through screenshots, clicks and keystrokes. How it works, why it's the most expensive and fragile way to reach an application, when it's still the right choice, and how to run it safely.Fri, 23 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/context-is-a-budget/https://www.agenticsystems.ai/blog/context-is-a-budget/An agent knows only what's in its context window, and every step re-reads all of it. How to spend that budget well: stable prefixes, just-in-time loading, notes, clearing and compaction.Wed, 23 Sep 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/designing-tools-for-agents/https://www.agenticsystems.ai/blog/designing-tools-for-agents/Tools are an interface for a new kind of user. Seven design rules, with a before-and-after on the same CRM data, and how to measure whether a tool change helped.Thu, 24 Sep 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/evals-for-agents/https://www.agenticsystems.ai/blog/evals-for-agents/An agent that succeeds once in five tries scores 100% on pass@5. Your users get every try, not the best one. How to measure agents honestly: repeated trials, pass^k, graders you trust, and the noise you didn't know was there.Fri, 02 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/five-workflow-patterns/https://www.agenticsystems.ai/blog/five-workflow-patterns/Most production "agents" are workflows: model calls arranged by code. Five patterns cover nearly all of them, and knowing them tells you when you really need an agent.Tue, 22 Sep 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/harness-engineering/https://www.agenticsystems.ai/blog/harness-engineering/As models got good enough to work for hours, the hard problems moved out of the prompt and into the software around the model. What a harness owns, seven principles from teams running agents at scale, and a small resumable harness you can crash on purpose.Thu, 01 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/humans-in-the-loop/https://www.agenticsystems.ai/blog/humans-in-the-loop/Asking a person to approve every agent action feels safe, but people approve almost everything they're shown. How to design approvals that stay meaningful: tier actions by risk, enforce the hard limits in code, show reviewers what matters, and escalate when an agent keeps pushing.Tue, 27 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/inside-a-coding-agent/https://www.agenticsystems.ai/blog/inside-a-coding-agent/Coding agents are the most widely used agents today, and the clearest example of everything in this series working together. What they actually do step by step, why their tools look the way they do, and why passing the tests is not the same as being done.Fri, 16 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/lethal-trifecta/https://www.agenticsystems.ai/blog/lethal-trifecta/Private data, untrusted content, and a way to send data out: any agent that combines all three can be turned against its user. Why model-level defences aren't enough, and six design patterns that make the attack structurally impossible.Wed, 30 Sep 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/mcp-explained/https://www.agenticsystems.ai/blog/mcp-explained/The Model Context Protocol in one post: what it standardizes, how a request travels, and why the July 2026 revision dropped sessions entirely. With a from-scratch server you can run as several instances.Fri, 25 Sep 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/memory-that-learns/https://www.agenticsystems.ai/blog/memory-that-learns/An agent that makes the same mistake on every task isn't learning. How agents keep lessons across tasks without retraining: what to write down, when, and how to stop memory going bad.Tue, 29 Sep 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/one-agent-or-many/https://www.agenticsystems.ai/blog/one-agent-or-many/Two well-known 2025 essays seemed to disagree about multi-agent systems. Read alongside a 260-configuration study and a 16-agent compiler build, they agree more than they disagree: it depends on the shape of the task.Sun, 27 Sep 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/prompt-caching-properly/https://www.agenticsystems.ai/blog/prompt-caching-properly/Agents re-read their whole context on every step, so prompt caching decides most of the bill. How the cache actually works, the setup that suits agent loops, the silent mistakes that switch it off (one of which costs more than not caching at all), and how to measure it.Fri, 30 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/rolling-out-agents/https://www.agenticsystems.ai/blog/rolling-out-agents/An agent that passed its evals can still surprise you on real work. How to roll one out in stages: shadow mode that compares its decisions with people's, written promotion gates, a canary on a small slice of real work, tripwires that watch behaviour, and a kill switch checked before every action.Fri, 06 Nov 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/routing-for-agent-cost/https://www.agenticsystems.ai/blog/routing-for-agent-cost/Most model calls inside an agent are small, narrow and repetitive, so they don't all need the biggest model. Four ways to cut cost, in the order to try them, and a harness that measures accuracy and dollars together.Tue, 06 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/should-this-be-an-agent/https://www.agenticsystems.ai/blog/should-this-be-an-agent/Most tasks that get pitched as agents shouldn't be. Six questions that tell you whether you need a single call, a workflow, a workflow with one agentic step, or a full agent, with worked examples and an interactive version for your own use case.Tue, 20 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/stop-calling-tools-start-writing-code/https://www.agenticsystems.ai/blog/stop-calling-tools-start-writing-code/Letting a model write a program that calls its tools can collapse dozens of round trips into one, but it isn't always cheaper. We measured both ways, found where the lines cross, and why the answer depends on how much data flows through the model.Sat, 26 Sep 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/the-agent-loop/https://www.agenticsystems.ai/blog/the-agent-loop/Every AI agent, from a weekend script to a coding agent that runs for hours, is built on the same loop. Here it is in 100 lines, with no framework.Mon, 21 Sep 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/time-horizons/https://www.agenticsystems.ai/blog/time-horizons/METR's 'time horizon' has become the best-known chart in AI. What it measures, what it doesn't, how fast it's moving, and how to estimate a horizon for your own agent, with code that runs without an API key.Sat, 03 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/watching-the-agents/https://www.agenticsystems.ai/blog/watching-the-agents/When an agent acts, someone needs to be able to answer: which agent was it, on whose behalf, what did it do, and what did it cost? Agent identity, tracing with OpenTelemetry's GenAI conventions, and the governance idea that started it all.Mon, 05 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/what-agents-cost/https://www.agenticsystems.ai/blog/what-agents-cost/Agent bills surprise people because an agent re-reads everything it knows on every step. A simple cost model, checked against a real run to within half a percent, the five levers that move the bill, and a calculator for your own workload.Tue, 13 Oct 2026 00:00:00 GMThttps://www.agenticsystems.ai/blog/why-multi-agent-systems-fail/https://www.agenticsystems.ai/blog/why-multi-agent-systems-fail/A study of 1,600+ annotated traces found 14 distinct ways multi-agent systems fail, and most of them aren't the model's fault. A tour of the taxonomy, what happened when we tried to reproduce three of them against a current model, and the design fixes that held up.Mon, 28 Sep 2026 00:00:00 GMT