For people running Claude Code who wonder whether a small classifier can send their prompts to cheaper models. Short answer: a little, with caveats. The code lives in my private infra repo; the numbers below are from its log.

My timeline has been full of “what Jev can do for you” posts. Jev sizes your request, Jev picks the right model, Jev saves you money. Fine. But every one of those setups has the same shape: before any work happens, your prompt goes to someone else’s model, just so that model can answer with one word.

Written for developers running coding agents on a real codebase. The implementation lives in lage; the operating notes are in .claude/memory-scoping.md.

I give my coding agents a memory. Mine is beads, where bd remember stores a durable note and a SessionStart hook replays them into every new session. After a year of use I had 51 notes: which endpoint nests items inside event, why the Fahrradmessstellen CSV is a rolling two-day window, that the Bonn council’s meeting links are not unique per meeting and keying a Svelte {#each} on one blanks the live site.

I spent a couple of evenings researching the b4mad.industries proposal for a Standard and Criteria Catalog for Fair, Transparent, and Sustainable Cost Sharing in FOSS Components and Dependencies. Going in, I assumed the hard part would be choosing between competing definitions of “fair.” Coming out, I’m convinced the more interesting finding is something else entirely: the FOSS assurance stack has a missing rung, and nobody is standing on it.

What I Found

We already have a tall stack of FOSS standards. SPDX/REUSE tell you what is in the software. ISO/IEC 5230 and 18974 (OpenChain) tell you whether the consuming organisation has a credible compliance and security program. CHAOSS measures project health. OpenSSF Criticality Score and Census II rank how important a dependency is. Tidelift, Open Collective, GitHub Sponsors, Sovereign Tech Fund, NLnet, Drips, Gitcoin Quadratic Funding, Optimism RetroPGF, ecosyste.ms Funds, Open Source Pledge all move money. Across all of these, not one defines a fair share, a disclosure schema, or a binding between dependency identifiers and funding flows [Source 1, 6, 7]. CHAOSS is closest — they have a Funding Working Group with a 2025 Practitioner Guide — but the guide explicitly states that “no single universal framework exists” and recommends customised approaches per funder.

As AI agent capabilities scale rapidly, the limiting factor for broad adoption is no longer model intelligence — it is security. Lex Fridman crystallized this in his widely-shared analysis: “security will become THE bottleneck for effectiveness and usefulness of AI agents.” This paper argues that the agent security problem is the primary differentiator in the emerging agent ecosystem, not model quality. We present the access–risk–usefulness triangle as a framework for reasoning about agent deployment, analyze why the current “YOLO mode” of agent usage cannot scale, and describe #B4mad’s architecture as a concrete, working implementation of security-first agent design.

Abstract

The vision of million-agent networks is compelling, but how do you actually build the discovery infrastructure to make it real? This article bridges the gap between theory and implementation, exploring practical patterns emerging from registry experiments, the Model Context Protocol (MCP) revolution, and production deployments.

We’ll examine four concrete approaches: DNS-based discovery, registry APIs, well-known URLs, and dynamic tool discovery through MCP. You’ll see how MCP acts as the “USB-C port for AI applications,” enabling runtime capability enumeration without hardcoded integrations. We’ll also tackle critical production challenges: the multiple context problem that fragments agent memory, security patterns for enterprise deployment, and the architectural decisions that determine whether your agent network scales or stalls at 1,000 agents.

A million AI agents collaborating, discovering each other, composing capabilities. We’re nowhere near. Today’s agents are integrated by hand, one at a time, on whatever protocol the vendor picked that week.

This is a discovery problem. We solved it for websites with DNS, for microservices with service meshes. For agents, we haven’t. Whoever does owns the next layer: an Agent Registration System, an Agent Naming Service, an Agent Gateway.

The interesting threshold sits at 10,000 agents. Below it, networks behave linearly and you can muscle through with manual integration. Above it, they self-organise. GPT Store crossed that line in January 2024 and growth went exponential — same platform, different physics.