标签：#Agents

Picture a moment every person who runs an AI agent knows. A task is halfway done and starting to go wrong. The agent took a weird turn a few steps back and now it is confidently heading somewhere bad. You have to decide fast on this. Do you step in? And if you do a quick "wait, check your work" nudge will that actually fix it? Or do nothing? Or worse knock a run that was about to recover on its own off the rails? That question is the whole project. Here is the honest short version of what I found. Detecting a failure is not fixing it A lot of recent agent research is about failure attribution — figuring out which step in a long run broke everything. Useful but it stops one step short of what you need when you are on call. Knowing where it broke is not the same as knowing what to do about it . So I asked a blunter question: given a failure, which fix actually recovers the run and which ones quietly make it worse? To answer it without fooling myself I rewind each failing run to the exact step where it went wrong, apply one fix, let it play forward and check the real answer against a hard ground truth no LLM grading another LLM. And I always compare against a "do nothing" control, because some runs recover on their own, and I did not want to give my fixes credit for that (or miss a "fix" that's actually worse than leaving the agent alone). What a capable agent actually gets wrong First surprise: a decent agent mostly doesn't fail in the dramatic ways people worry about. It rarely loops, rarely forgets to answer, rarely fumbles a tool that throws an error in its face. It fails in two quieter ways and both are the same underlying mistake: acting on the surface of the situation instead of the real thing underneath. It makes up an answer it could have looked up. The fact it needs is sitting right there behind a tool call it just never makes, so it fills the gap with something plausible. Reads "manager: #202," never looks up who #202 is, asserts a name anyway. It trusts a t

A voice agent can sound polished, respond instantly, and still create a trust incident in one sentence: “Stop calling me.” If that request only updates the SMS path, your agent may keep dialing tomorrow. If it only updates a call transcript, your follow-up workflow may keep texting. For builders shipping AI callers, inbox agents, scheduling bots, or multi-step outreach workflows, consent is no longer a static checkbox. It is runtime state. That is where an AI consent ledger helps. It gives every agent action a simple rule: before contacting, enriching, recording, or escalating a person, check the latest consent state from one durable place. This guide shows how to design that ledger without turning your product into a compliance maze. This is technical architecture guidance, not legal advice. If your workflow touches regulated outreach, health, finance, employment, or sensitive personal data, involve a qualified legal reviewer. Why AI agents make consent harder Traditional apps usually ask for permission at predictable moments: signup, newsletter opt-in, cookie banner, phone number capture, or billing consent. AI agents blur that boundary. A production agent may: answer an inbound call summarize a voicemail text a follow-up link schedule another call enrich a CRM record trigger a campaign step next week hand the case to a human retry after a failed tool call switch from voice to SMS or email Each step may be valid by itself. The risk appears when consent changes in one channel and the rest of the workflow does not notice. The common failure shape is simple: User revokes permission in the channel in front of them. The agent logs the message as conversation text. Another workflow keeps running because it never checked revocation state. That is not an LLM problem. It is a state-management problem. What is an AI consent ledger? An AI consent ledger is an append-only record of permission events plus a fast read model that answers one question: Is this specific agent allo

Handoffs are useful when a specialist agent needs to take over a task. They also make cost easier to hide, because the bill is spread across graph nodes instead of one visible chat turn. Why can LangGraph handoffs multiply tokens? LangGraph handoffs can multiply tokens because each model-calling node may resend instructions, prior messages, retrieved material, tool returns, summaries, and artifacts, then loops or handoffs repeat that payload for the next agent. Token amplification is the total prompt-plus-completion tokens across a trace divided by a simpler baseline for the same task; Anthropic reported in June 2025 that multi-agent systems used about 15x more tokens than chats while improving an internal research evaluation by 90.2% . Quick Answer: Handoffs raise the token bill when each agent receives copied context instead of a narrow task packet. Anthropic’s June 2025 research system showed the tradeoff clearly: multi-agent runs used about 15x more tokens than chats while scoring 90.2% higher on its internal research evaluation . In LangGraph, the practical issue is observability and budgeting, not whether graphs are bad. The LangGraph project describes the runtime as a way to build stateful, long-running agents with persistence, human control, memory, and debugging support; those same traits make it possible to measure where context grows instead of guessing. "Multi-agent systems are often highly effective at open-ended research tasks, but token usage can be substantial," — Anthropic engineering team at Anthropic The small verified demo below shows the arithmetic behind a 15x bill: a 100-token task becomes 1,500 billed tokens when 5 agents each receive 3 copies of the relevant context . """ Tiny token-accounting demo: handoffs multiply the same task context. """ task_tokens = 100 agents = 5 context_copies_per_handoff = 3 # instructions + task + summary/history direct_bill = task_tokens handoff_bill = task_tokens * agents * context_copies_per_handoff print ( f

I've spent enough time in the trenches of software engineering to know that there is nothing more soul-crushing than writing 'glue code.' You know exactly what I mean—the thousands of lines of boilerplate, error handling, and webhook listeners required just to make two services talk to each other. When Bland AI first arrived on the scene, it was essentially another API you had to integrate. You'd write a Node script, handle the async nature of outbound calls, manage your credentials in environment variables, and then spend weeks building a dashboard just so you could see what happened during a call. It worked, but it wasn't intelligent. The shift we are seeing right now with the Model Context Protocol (MCP) changes the fundamental architecture of integration. We are moving from 'integration as an engineering task' to 'integration as a capability.' Instead of writing code to bridge Bland AI and your application, you provide an MCP server that gives your LLM—whether it's Claude or Cursor—direct access to those telephony tools. I recently started using the Bland AI MCP server via Vinkius, and the difference in how I can orchestrate workflows is night and day. This isn't about just 'making a call.' It's about giving an agentic loop control over a communication channel. The Architecture of Voice Orchestration When you look at traditional API integrations for something like Bland AI, you focus on the request/response cycle. You send a payload to trigger a call, and then you wait for a webhook to notify your backend that the call is finished. With this MCP server, the mental model shifts. You aren't managing webhooks; you are managing tools. The toolset provided here—including send_phone_call , create_voice_agent , and list_recent_calls —allows an LLM to act as a telephony engineer. Here is what happens when you actually use it in Cursor or Claude: You don't just say "Make a call." You can instruct the agent, "Look at my recent calls from yesterday, find any where the tran

TL;DR: MCP servers run with significant privileges inside AI agent pipelines, and most teams ship them without any security review. mcp-security-scan is an open-source CLI and GitHub Action that checks for credential theft patterns, data exfiltration, unsafe execution, and code obfuscation — and outputs a 0-100 trust score that integrates with AgentGraph's identity layer. The Moltbook breach last year is still the clearest example of what happens when you scale agent infrastructure without thinking about trust. 770,000 agents, zero identity verification, and when it went down it exposed 35,000 emails and 1.5 million API tokens. The tokens were the real problem — many of them were credentials passed through MCP servers that nobody had audited. MCP (Model Context Protocol) servers are the connective tissue of modern agent systems. They sit between your LLM and the outside world, handling tool calls, filesystem access, API requests. That position gives them a lot of power. It also makes them an obvious target. And yet most teams treat MCP servers like they treat npm packages circa 2015: install and trust. What Actually Goes Wrong Before getting into the scanner, it's worth being specific about the threat categories. There are four that show up most often in real codebases: Credential theft — MCP servers that read environment variables indiscriminately, log request/response payloads, or forward tool call arguments to external endpoints. This one is subtle because the server might be doing legitimate work and exfiltrating credentials. Data exfiltration — Outbound HTTP calls to domains that weren't declared in the server's manifest, or calls that happen inside tool handlers where the LLM can influence the destination URL. Prompt injection into tool parameters is the attack vector here. Unsafe execution — eval() , exec() , subprocess calls, or dynamic require() / import() where the argument comes from tool call input. If an LLM can influence what gets executed, you have a

As Meta pours billions into AI infrastructure and agents, Zuckerberg is working to convince investors that the payoff will be worth the price.

In my previous article , I built a fully local RAG assistant Ollama, ChromaDB, LangChain, all running in Docker. It answered technical support questions by searching through documentation and citing sources. It worked. But after using it for a while, I noticed something uncomfortable: it treated every question the same way . Ask it "how to close monthly payroll?" it searches the docs. Fine. Ask it "the server crashes at startup" it also searches the docs. Less fine. Ask it something completely outside the documentation it searches the docs. Useless. A real support technician doesn't do that. They first assess the situation, then decide what to do: look it up, run a diagnosis, or escalate to a human. My RAG had no such judgment. That's what this article is about how I evolved the system into an Agentic AI architecture using LangGraph, where the assistant first decides which strategy to use , then acts accordingly. The Core Limitation of Classic RAG Classic RAG is a linear pipeline. Every query follows the exact same path: Question → Embed → Retrieve → Prompt → LLM → Answer No branching. No decision-making. No memory between steps. This works perfectly for procedural questions where the answer lives in the docs. But technical support involves at least three distinct scenarios: Scenario Example Best strategy Procedural question "How do I create an account?" Search documentation Known error code "ERR-COMP-001 appears" Lookup error database Unknown incident "Server crashes, no idea why" Diagnose + escalate if needed A single RAG pipeline handles the first case well and the other two poorly. The solution is to add a layer of reasoning before retrieval. What Agentic AI Adds The shift from RAG to Agentic AI comes down to one thing: the system plans before it acts . Instead of one fixed pipeline, you have: Question ↓ Classifier (what kind of question is this?) ↓ ├── Procedural → RAG Agent (search docs) ├── Error code → Diagnostic Agent (lookup + LLM analysis) └── Complex → D

A few weeks ago I wrote up what agents were paying for on NetIntel , my platform of pay-per-call APIs settled in USDC over x402 — no signup, no API keys, no accounts. An agent hits an endpoint, gets a 402 Payment Required , pays a fraction of a cent, and gets structured data back. That's the whole loop. Since then the dataset has grown, I've instrumented every settled call into a proper database (payer wallet, endpoint, price, latency, transaction hash), and I've launched a second settlement rail. So this is the rewrite with real numbers instead of eyeballed ones — and the findings didn't soften. They sharpened. The setup 2,646 settled paid calls from 194 distinct paying wallets, across 101 live endpoints , over six weeks of instrumented production data. Every call in this dataset is a real on-chain payment with a transaction hash — no test traffic, no estimates. Settlement runs on Base, and as of this month on Solana too. This is still one platform's data in a young ecosystem — the caveats are at the bottom, and one of them is bigger than it looks. But the shape has now held for six weeks straight, and it's the same shape I flagged the first time. Finding 1: revenue is absurdly concentrated — and it stayed that way Five endpoints drive 69% of all revenue. One of them — a text-to-structure endpoint that takes messy input and returns strict typed JSON — is 42% by itself . The rest of the top five are all in the same family: translation, structured LLM inference, and one domain-intelligence report. The other 96 endpoints split the remaining 31%. Thirty-five of the 101 have never been paid for once. Not "underperformed" — zero settled calls, ever. When I first published this pattern I wondered if it was an artifact of a small sample. The dataset has since more than doubled and the concentration ratio barely moved. I now treat it as the market talking, not noise. Here's the part I'd want to know if I were reading this: that 42% endpoint is essentially one buyer — a wall

AI agents often begin as application code: a prompt, a model call, a few tools, and enough control flow to make the first example run. As the agent grows, the definition and the environment tend to collapse into each other. Model configuration lives beside credentials. Tool access is mixed with prompts. Persistence assumes a particular process. Deployment choices become part of the agent itself. That coupling makes an agent harder to inspect, test, move, and recover. We built Clear Ideas Agent Runtime around a different boundary: the agent definition should be portable, while the infrastructure that executes it should remain under host control. The Agent Manifest is the portable contract An Agent Manifest is a versioned YAML or TypeScript definition. It can describe: prompts and structured outputs; typed variables; conditions and loops; tools and MCP connections; approvals and webhooks; sandboxed code steps; sub-runs; limits and final outputs. A separate Agent Run Manifest supplies the values and execution choices for one invocation. That keeps the reusable agent definition distinct from the inputs and operational decisions associated with a particular run. Here is a small example: schemaVersion : " 1.0" name : research-brief variables : topic : type : string researchNotes : type : string briefDraft : type : string steps : - id : research type : prompt prompt : | Research {{ topic }} and return concise notes. outputVariable : researchNotes - id : draft type : prompt prompt : | Draft a brief using these notes: {{ researchNotes }} outputVariable : briefDraft The manifest describes the agent. It does not contain the credentials, infrastructure account, or persistence implementation that happens to run it. The host controls the operational boundary The host application supplies and controls: models and provider credentials; connections and tool authorization; persistence and artifact stores; local or remote compute; sandbox providers; telemetry; concurrency and resource

Microsoft has released a reference architecture for routing agent traffic on Azure Kubernetes Service. It breaks down the issue into three key choices: which model answers a call, how the call is managed, and which GPU replica handles it. By Claudio Masolo

An AI agent can turn one user request into a small storm of model calls, MCP tool calls, retries, partial failures, and background work. If you only meter the final response, you are guessing. If you meter every low-level event without context, you create noise customers will not trust. That is the billing trap many AI product builders are walking into: the product feels simple, but the usage behind it is multi-step, non-deterministic, and easy to dispute. MCP makes this more urgent. The Model Context Protocol gives agents a standard way to call tools, but a standard tool call is not the same thing as a fair usage meter. A production meter needs to answer harder questions: Which customer, workspace, user, and agent run caused the call? Was it read-only or write-capable? Was the call retried, duplicated, cached, rejected, or actually executed? Did it hit a paid upstream API? Should it count toward quota, invoice, abuse limits, or only observability? Can you explain the charge without exposing private prompt or customer data? This guide shows a practical MCP usage metering architecture for solo developers, micro product teams, and AI platform builders who need cost control without surprising users. Why MCP Usage Metering Is Different From Token Tracking Token tracking is mostly linear. You send a prompt, receive a response, and record input tokens, output tokens, model, latency, and cost. Agent tool usage is messier. A single request like "research these accounts and update the CRM" might trigger: A retrieval call to fetch customer rules A search tool call for each account A browser or enrichment call for missing fields A CRM read A CRM write proposal A human approval pause A final write call A summary response Some calls are internal. Some are customer-visible. Some are expensive. Some are dangerous. Some are free but should be rate limited. Some fail after doing real work. Some are retried by the agent, the SDK, the queue, or the network layer. If you charge blindly

I spend a lot of time in the AI space -- reading papers, building things, talking to engineers who are actually shipping. And there is a gap between what the demos show and what production systems actually look like that nobody is being fully honest about. So here is my honest take on where things actually are. The Problem With How We Talk About AI Agents Everyone is calling everything an "agent" right now. A function that calls a tool? Agent. A chatbot with memory? Agent. A script with a loop? Agent. This dilution is not just semantic. It is causing real engineering mistakes. When you do not have a precise definition for what you are building, you end up over-engineering simple pipelines and under-engineering genuinely complex ones. I have seen teams spend weeks adding "agentic" orchestration to workflows that would have been fine as a single well-structured prompt. Here is the definition I keep coming back to: an agent is a system that has an objective, not just an instruction. It decides what to do next. It handles failure. It knows when it is done. Everything else is just a fancy function call. 🟢 If your system needs a human to tell it each step, it is not an agent. It is a chat interface. 🔵 If your system can recover from a failed tool call and try a different approach, you are getting somewhere. ✅ If your system can decompose a goal into subtasks and delegate them, that is the real thing. What Is Actually Happening in Production Right Now The honest picture from teams I follow and talk to: Most real agent deployments are narrow. They do one thing well. Customer support triage. Document extraction. Code review on a specific codebase. They are not general-purpose reasoning engines. They are purpose-built pipelines with some intelligence in the decision layer. The teams getting good results are not chasing the latest model release. They are obsessing over: ☑️ Tool design -- what can the agent actually call, and how clean is the interface ☑️ Failure handling -- wh

I spend a lot of time in the AI space -- reading papers, building things, talking to engineers who are actually shipping. And there is a gap between what the demos show and what production systems actually look like that nobody is being fully honest about. So here is my honest take on where things actually are. The Problem With How We Talk About AI Agents Everyone is calling everything an "agent" right now. A function that calls a tool? Agent. A chatbot with memory? Agent. A script with a loop? Agent. This dilution is not just semantic. It is causing real engineering mistakes. When you do not have a precise definition for what you are building, you end up over-engineering simple pipelines and under-engineering genuinely complex ones. I have seen teams spend weeks adding "agentic" orchestration to workflows that would have been fine as a single well-structured prompt. Here is the definition I keep coming back to: an agent is a system that has an objective, not just an instruction. It decides what to do next. It handles failure. It knows when it is done. Everything else is just a fancy function call. 🟢 If your system needs a human to tell it each step, it is not an agent. It is a chat interface. 🔵 If your system can recover from a failed tool call and try a different approach, you are getting somewhere. ✅ If your system can decompose a goal into subtasks and delegate them, that is the real thing. What Is Actually Happening in Production Right Now The honest picture from teams I follow and talk to: Most real agent deployments are narrow. They do one thing well. Customer support triage. Document extraction. Code review on a specific codebase. They are not general-purpose reasoning engines. They are purpose-built pipelines with some intelligence in the decision layer. The teams getting good results are not chasing the latest model release. They are obsessing over: ☑️ Tool design -- what can the agent actually call, and how clean is the interface ☑️ Failure handling -- wh

Docker published the second entry in its Coding Agent Horror Stories series on July 20, and the operational read is short: on a stock developer laptop, an AI coding agent runs with the engineer's filesystem permissions and the engineer's credentials, with nothing sitting between it and the host. The post frames a scenario in which the agent deletes production and works backward through why that outcome is not exceptional. Docker names the piece as part two of a series that will cover six categories of coding-agent failure. What the post actually claims Two claims carry the argument. First, the agent inherits the developer's shell posture: whatever the developer can touch on disk, the agent can touch; whatever token is exported into the environment, the agent can spend. Second, that default is not a sandbox. Docker's phrasing is that nothing sits between the agent and the host unless the operator puts it there. The piece does not attribute the scenario to a named incident; it is a category, not a case study. Anyone extrapolating specific companies, victims or numbers is filling in blanks the source did not. The runner problem, one hop to the left For CI operators this shape is familiar. A self-hosted Actions runner or a Jenkins agent that mounts the workspace, holds a checkout token and can call the host shell is a service you already isolate on purpose. You isolate it because the workflow you invited in is not always the workflow that runs. You isolate it because the token in the environment can do more than the job description. You isolate it because rollback of a bounded container is cheaper than reasoning about everything a process touched on a shared box. A coding agent living on the developer laptop occupies the same trust position, one machine earlier in the pipeline. It reads and writes the working tree. It holds session credentials to cloud APIs, the cluster and the registry. It executes instructions the developer did not always write, sometimes routed from

This article provides a step-by-step guide to building and testing a cross-cloud currency agent. A coordinator built with Strands Agents and hosted on Amazon Bedrock AgentCore Runtime (in AWS us-east-1 ) discovers and delegates to a Google ADK agent (on GCP Cloud Run in us-central1 ) over A2A v1.0 , cross-checks results against an MCP exchange-rate tool , and measures what independent cross-cloud verification costs in latency, reliability, and overhead. What is This Project Trying to Do? Most Agent-to-Agent (A2A) protocol demos stop at "look, the HTTP 200 OK request succeeded." That is a smoke test, not an interoperability benchmark. This project goes further: an Amazon Bedrock AgentCore-hosted Strands Agents coordinator discovers and delegates to a Google ADK agent running on GCP Cloud Run, comparing the results against a local MCP stdio exchange-rate tool backed by live Frankfurter daily reference rates. We also compare the performance, developer experience, and wire compatibility directly against our previous benchmark run hosted on Microsoft Foundry in Azure ( gpt-5-mini ), giving us a true cross-cloud benchmark across AWS, Azure, and GCP. The questions we answer with hard empirical data rather than vibes: Can an AgentCore-hosted Strands agent discover and invoke a Google ADK agent through an A2A agent card with no framework-specific glue? What latency and token overhead does remote-agent verification add? Does independently verifying an MCP tool result over A2A improve correctness or failure recovery enough to justify that overhead? How does AWS Bedrock AgentCore Runtime compare like-for-like with Microsoft Foundry on Azure? Reduce, Re-Use, Re-Cycle! This builds directly on the currency agent from the previous articles in this series: Getting Started with MCP, ADK and A2A | Google Codelabs GitHub - jackwotherspoon/currency-agent That agent — built with Google ADK, Gemini 2.5 Flash, and a FastMCP exchange-rate server backed by the free Frankfurter API — serves a

How agents remember - and why deciding what to forget is the real skill An agent that starts every step with a blank mind cannot really pursue a goal. It would reintroduce itself to you on every message, forget what it just tried, and repeat the same mistake forever. Memory is what turns a stateless model into something that accumulates - that knows who you are, what it has already done, and what it learned last Tuesday. This post is about how that works and, more importantly, about the discipline of deciding what an agent should remember at all. The context window is not memory. The first thing to unlearn: a model’s context window is not its memory. The context window is working memory - RAM, not a hard drive. It is finite, it is reset on every request, and every token in it costs money and dilutes the model’s attention. Stuffing an entire conversation history and knowledge base into the prompt does not scale, and past a point it actively hurts - the model loses the important signal in a sea of stale detail. Real memory lives outside the window and is selectively loaded into it when needed. Four kinds of memory Borrowing loosely from cognitive science, agent memory is usually split into four types, and good systems use all of them: Short-term/working memory - the current conversation and the agent’s recent thoughts and observations. Lives in the context window. Long-term episodic memory - a record of what happened : past conversations, decisions, and the outcomes of previous tasks. Long-term semantic memory - facts and knowledge: who the user is, domain information, documents. This is what retrieval-augmented generation pulls from. Procedural memory - how to do things : learned skills, tool-use patterns, and reusable strategies. Short-term memory: the rolling buffer The simplest memory is just keeping recent turns in the prompt. The problem is that conversations outgrow the window, so the standard move is to keep the last few turns verbatim and summarise the older

Hello, I'm Rijul. I'm building git-lrc, a micro AI code reviewer that runs on every commit. It's free and source-available on GitHub. Star git-lrc to help more developers discover the project. Do give it a try and share your feedback Past few years, when things like GPT came out, we were mainly learning how to use them. Which basically meant using better prompts. So we were continuously refining prompts and trying to get better outputs. And that is fine. This was basically an exchange back and forth between us and the bot. We ask something. The AI responds. We refine our prompt. The AI responds again. This loop continues. Agents coming into the picture But now we are seeing the rise of AI agents. We see them everywhere, in every LinkedIn post, every YouTube video. The main idea behind agents is to hand over our work to them, and they will handle the remaining tasks with less constant back and forth compared to traditional chatbots. Here, agents are expected to work in loops. They do one thing. Then they review the result, identify additional things that need to be done, take action again, and repeat. The main input we give is usually a goal or a prompt. But this alone will not give us better control over how the agent works. Reliable agents also need things like success criteria, access to the right tools, feedback mechanisms, and rules for when to stop or ask for human input. The agent works in loops. So we need a better way to design and control these loops. Introducing Loop Engineering Agents perform tasks through loops. So instead of only focusing on prompt engineering, we also need to engineer the loop in which the agent operates. This is where Loop Engineering comes in. The term has recently gained attention as people started exploring better ways to build reliable workflows around AI agents. The idea is simple: Instead of only telling an agent what to do, we design the process that helps the agent complete the task reliably. Prompting is still an important pa

So I had this idea that kept nagging at me. Every AI chat app works the same way. You type something, the model returns text or markdown, the UI renders it as a nice formatted paragraph. That is fine if you want an answer. It is genuinely boring if you want to actually build something. What if the AI could respond with a working game board you could click? What if saying "make it Barbie themed" actually transformed the whole interface while you watched? What if "add a starfield in the background" dropped an animated canvas behind your chat in real time? I spent a few weeks building exactly that. I call it FlowChat . Here is the live version: https://flowchat-public.varshithvh.workers.dev And yes, someone immediately asked it to play Tic Tac Toe and then asked it to switch to an Oppenheimer theme mid-game. I could not be prouder. The Idea Normal AI chat: model returns markdown, client renders it as text. Simple, predictable, boring. FlowChat: model returns raw HTML with CSS and JavaScript, client injects it directly into the DOM using a streaming protocol built on the browser's native template system. That one change makes the entire experience different. You are not reading about a game. You are playing one. You are not reading about a Barbie color palette. You are sitting inside one. The AI does not just answer questions. It rebuilds the UI from its responses . What You Can Actually Do With It I want to give you a feel for what this means in practice before getting into the technical bits, because the demos are more interesting than any architecture diagram. Games : Ask it to build Tic Tac Toe. You get a playable board, click-to-move, an AI opponent, win detection. Ask for Connect 4. Ask for Snake. The game renders in the chat as an agent bubble with a form inside it. Each move submits to the LLM which processes it and updates only the cells that changed. Themes : Say "change to a Barbie theme". The model injects CSS overrides and the whole interface turns pink. Me

Ten agent sessions ("minds," in this codebase) run continuously on one box, each with its own responsibility — one writes code, one talks to me on Telegram, one watches sensors, one just measures the fleet itself. They coordinate the way a lot of multi-agent systems eventually do: a shared log file, one line per event, [task] / [taking] / [done] . That log is fine for "what happened." It is useless for "what do we owe, and how much did it cost" — the two questions I actually needed answered before I was willing to let the fleet run unattended overnight. The board is not a ledger, but it can feed one The fix wasn't a new coordination protocol. It was noticing that every line on that board is already a transaction if you're willing to look at it that way: board event ledger meaning [task] fix-the-thing a liability opens [taking] pub: fix-the-thing the liability moves to a specific debtor [done] pub: fix-the-thing the liability settles a provider round-trip (one agent turn) a unit of labour is spent So the board gets replayed into three separate double-entry hledger journals, each tracking a different commodity: money — imputed USD (token counts priced through one rate table). promises — commodity PROMISE : an open [task] with no matching [done] is a standing liability, not a line that scrolled off screen. labour — commodity TURN : one provider round-trip, the fungible unit every mind actually spends, regardless of whether it's writing code or answering a sensor. Each journal gets checked two independent ways — hledger check for internal parity, plus a second, independently-written replay of the same board that has to agree with the balance query. A booking bug fails loud, not silently, because two things that should compute the same number just disagreed. Querying "who owes what" stops being a grep and starts being a query: $ mesh-promises --balance standing open obligations (bal liabilities:promises · 1 PROMISE = open, netted): 1 PROMISE liabilities:promises:pub:chat

In my previous article, I argued that AI changes the role of constraints in software...