$ today_in_10_seconds
Three Show HN projects landed today that all assume the same thing: the agent is the primary user, not the human. A self-hosted runtime treats audit and RBAC as runtime primitives, a headless CRM exposes itself as grepable plain text for agents to drive, and a Rust agent literally rewrites its own binary. The interface layer for software is quietly being rebuilt for machines.
$ lead_story
Democr.ai — self-hosted agentic runtime with audit and RBAC
A new Python framework, Democr.ai, bundles the messy operational concerns of agentic apps — audit, RBAC, OS-level sandboxing, multi-tenancy, model orchestration — into a single runtime contract rather than leaving them as things you bolt on later. Its distinctive bet is server-driven UI built on Google's A2UI protocol, so agents compose native application surfaces declaratively instead of returning markup or screenshots. Everything, including auth, is a module built against one SDK boundary, and its threat model is explicit: it defends against accidental supply-chain compromise, not a deliberately malicious module you chose to install. For teams in regulated environments, this is the increasingly common pattern — treat observability and audit as day-one primitives, not instrumentation added after the incident.
// why_it_matters: Agentic infrastructure is converging on 'audit and isolation are runtime primitives,' not app-layer conventions.
$ also_worth_attention
crmkit — an agent-first CRM driven over plain HTTP
crmkit is headless by design: no UI, no SDK, just a plain-text grepable API and a one-page manual the agent loads at runtime. Records return one labeled line each (JSON on demand), errors ship a hint field so the agent self-corrects without a schema, and it speaks MCP at /mcp for one-click connection to Claude or ChatGPT. The insight worth stealing: plain text is token-cheap and survives context truncation — a real design constraint when your primary user has a context window.
link →Wizard — a self-extending autonomous agent in one Rust binary
Wizard is a single Rust TUI agent that extends itself at runtime: new skills, MCP servers, and subagents are plain files, and /evolve --deep will propose a diff to its own source, rebuild, smoke-test, and swap its binary — with the old one kept for a one-command rollback. It argues self-extension is a security feature, since every install converges on a different tool surface, so no single exploit transfers. The honest caveat is right in the README: no sandbox, auto-approved tool calls, run it in a VM.
link →goose goes vendor-neutral under the Linux Foundation
Surfacing from Wizard's own comparison table: Block's goose agent is now governed by the Linux Foundation's newly formed Agentic AI Foundation, making one of the deepest MCP ecosystems (70+ documented extensions) vendor-neutral. Neutral governance is how an agent framework becomes safe infrastructure to build a business on rather than a single vendor's bet.
link →
$ bigger_picture
The trend behind the headlines
Today's three projects rhyme. Each one accepts that the human is no longer the main operator — the agent is — and redesigns the contract accordingly. crmkit optimizes its API for token cost and context truncation; Democr.ai makes audit and permission runtime guarantees because you can't trust an autonomous caller by convention; Wizard hands the agent the keys to its own source code and then spends most of its README explaining the guardrails. That last point is the tell: as agents gain autonomy and self-modification, the interesting engineering shifts from capability to containment — sandboxing, rollback, audit trails, and explicit threat models. The builders shipping the durable stuff are the ones treating 'what can go wrong when nobody's watching' as the core feature.
Model Context Protocol (MCP)
All three of today's projects speak it, which is the real story. crmkit exposes its whole API through a single generic MCP request tool; Wizard registers MCP servers at runtime via /evolve; goose built one of the earliest and broadest MCP integrations. If you're building anything agent-facing, MCP is now the default connective tissue rather than an optional extra.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime