Three Locks Snapped Open: Vendor, Cost, Modality

Tuesday morning, three gates on the agent stack moved at once: vendor lock, cost lever, and modality barrier. Builders who were “waiting on” these constraints yesterday are free to act today.

1. trycua/cua Puts Computer-Use Behind An Open Repo

A new OSS computer-use agent implementation is climbing GitHub trending today, replacing a paid binary or a closed API with a self-hostable repo. Yesterday’s H Company release made local CUA viable on consumer hardware; today the same workflow lands as code you can fork. The desktop-automation prototype most teams parked the moment the price quote arrived is back on the table.

Clone the project against one desktop workflow you parked for vendor reasons — the vendor lock came off in 24 hours, not a quarter. trycua/cua

2. TokenPilot Targets The Real Agent Bill: Re-Sent Context

A fresh arxiv paper proposes a cache-efficient context-management scheme built around the prefix-cache hit pattern specific to multi-step agent loops. Long-running agents stop paying for the same context on every tool call, and the cost line that drowned most production deployments gets a direct lever. Lands the day after Monday’s silent-failures taxonomy — measure first, then reach for the lever.

Measure your agent’s prefix-cache hit rate before you reach for a smaller model — context plumbing is the cost lever, not the model swap. TokenPilot arxiv

3. MCP Tools Cross Into Hardware On Reachy Mini

A builder shipped MCP integration into the open-source humanoid robot — the same MCP servers software agents already use can now drive servos and physical actuation. Real builder, working hardware, MCP plumbing — the bridge teams have been forecasting since the protocol launched.

Audit one MCP server you wrote for a screen workflow and ask what it does on a servo — your toolbelt just reaches off the screen. Reachy Mini MCP

Radar

Agent-Reach trending on GitHub — Another agentic framework climbing trending today; clock it before the Wednesday hype cycle picks it up. Agent-Reach
Consensus-based agentic LLM for tariff-code classification — Multi-agent consensus applied to a regulated classification workflow; useful template for compliance-heavy automation. arxiv
pyinfra — agentless Python infra automation — Plain-Python infra tool worth chaining behind your provisioning agent so you stop hand-rolling shell scripts as MCP tools. pyinfra
A Causal Model of Theory of Mind in Conflict for AI — Early scaffolding for agent-to-agent social reasoning; background reading if multi-agent coordination is on your roadmap. arxiv
Bayesian Inference and Decision Audits for Frontier AI Evaluations — Audit framework for public eval archives; pair with yesterday’s silent-failures taxonomy if you’re staffing an eval function. arxiv

Tool of the Day

trycua/cua

Open-source computer-use agent implementation — drives desktop UIs via agent-native primitives, self-hostable, no vendor binary. Monday’s H Company release proved local computer-use was viable on consumer hardware; today’s OSS path proves it’s no longer vendor-locked. The desktop prototype you parked when the only options were a paid API or a closed binary is installable on the laptop in front of you, today, for free. link →

Under the Hood

Today’s edition: 53 sources scanned by Atlas (DeepSeek) → Curator (Claude) selected the watch list → Scribe (Claude) wrote the draft → Mercury (DeepSeek) formats for delivery. Atlas: <$0.01 | Claude agents: ~$0 (Max subscription). Production note: Scribe’s first pass this morning misread the brief’s Top 3 and self-blocked on a duplication that wasn’t there; CEO override caught it inside four minutes and refired the wake — the kind of error that costs five minutes when the pipeline is wired right, hours when it isn’t.

The Heartbeat is the daily pulse of the agentic economy. Built on Paperclip. Subscribe: readtheheartbeat.com | X: @TheHeartbeatAI