Source: raw/Don_t_build_more_AI_agents_until_you_watch_this.md — YouTube, Nate B Jones (“AI News & Strategy Daily”, video id BOXK2XFLA-E) · refreshed 2026-10-27 with raw/Ep._244_-_How_Alibaba.com_Is_Building_the_Future_of_Agent-to-Agent_Commerce.md (The Artificial Intelligence Show)
Nate B Jones argues the next phase of agent work is maintenance, not construction: the durable, ownable layer around an agent is its harness (a.k.a. workbench), and the discipline that matters in 2026 is pruning and auditing that harness rather than piling on more tools. The anchor claim — Vercel made its sales-inbox agent better by deleting ~80% of its tools — reframes “add a capability” as the beginner instinct and “ask what to remove” as the maintenance instinct. This complements (and partly contradicts) the wiki’s Canva case study, which concludes the harness is disposable and evals are the durable asset.
Key Takeaways
- The harness is the maintainable layer. Agent = the worker; harness = the workbench: what it reads, what it remembers, which tools it can touch, what it’s allowed to change, what proof it must bring back, and what stops it when work gets risky. The model moves underneath you; the harness is what you actually own and maintain.
- Tool-pruning over tool-piling. “The beginner instinct is to add; the maintenance instinct is to ask what should be removed.” Concrete anchor: Vercel deleted ~80% of its sales-inbox agent’s tools and it improved — built by studying one top rep’s observed workflow, not the documented one.
- Agents break in two directions — a falsifiable framing not previously in the wiki: (1) world-drift — sources, workflows, and definitions around the agent change; and (2) model-improvement — a tool that helped a weaker model can confuse a stronger one, and a guardrail that protected you from a clumsy model can trap a better one.
- Agents inherit the crud of the systems around them. A stale wiki page, a renamed CRM field, or a drifted dashboard definition is merely annoying to a human but dangerous to a proactive agent that produces convincing output from a bad source.
- Frontier labs are betting better models ship the harness faster. Codex and Claude Code are framed as carefully-maintained harnesses (terminal, browser, computer-use, plugins, memory, approvals, sandboxing, logs) — a compounding flywheel where the model improves the workbench that runs the model.
- The more custom the harness, the more you own the upkeep. Light harness = instructions + memory + source folders + repeatable methods + proof/escalation rules. Deep harness = data feed, review screen, permission levels, logs, model choice, escalation paths, approval rules — plus a plan for what happens when the model changes.
October 2026: “agent equals model times harness times context” (Alibaba.com)
From The Artificial Intelligence Show (Paul Roetzer) ep 244, an interview with Alibaba.com president Kuo Zhang (caption “Quo”) on Accio (caption “Axio”), its team of agents for small-business buyers and sellers (raw/Ep._244_-_How_Alibaba.com_Is_Building_the_Future_of_Agent-to-Agent_Commerce.md, https://www.youtube.com/watch?v=pTLdkUZVpEk):
- The formula. “Agent equals to model times harness times context.” It multiplies because if any element is zero, “the entire equation is zero.”
- Model: frontier models including Qwen, plus post-trained small Qwen models, with serving work for cache rate and performance.
- Harness: “all this logic, the loops, the guardrails, the tools,” including connectors to the many platforms sellers use (aggregate data, upload listings, market research).
- Context: “all the real business data, the conversations you have and also the data from platforms.”
- Measured, not assumed. Alibaba built its own “commerce agent bench” to track improvements. Zhang adds that six months earlier “I cannot answer that question” of what an agent is made of.
- How it fits Nate’s argument ^[inferred]: a model upgrade cannot rescue a zero in the harness or context term, which is why the audit below targets the harness rather than the model.
The 5-point agent-maintenance audit
Run this on any agent you already operate (not a build checklist — a maintenance checklist):
- What’s it eating? Are its sources still current? Did the workflow move? Did an old source quietly become misleading?
- Test its reach. Is it read-only, or can it draft, create tickets, post to Slack, update records, spend money, publish? Re-verify the blast radius.
- Check its job. Don’t let a summary agent silently become a planning agent. Change the job on purpose, not by drift.
- Check the proof. Demand a linkable trail a human can inspect — linked tickets, quoted customer language, an explicit list of sources it couldn’t access — not the agent’s own say-so.
- Check the value. Does anyone read its output? Does it save time after review? Should it be rebuilt or retired?
Related
- Inside Gamma’s Creative Agent Harness — a production harness: about 200 calls to 20 models per deck, vision QA, A/B tests over evals.
- Canva AI 2.0 case study — the productive tension: Canva concludes “harness is disposable, evals are durable”; Jones argues the harness is the maintained, ownable layer. Read together.
- Production Class Ladder — same author; the tool-pruning instinct rhymes with that article’s “a ladder needs a down-staircase.”
- 12-Factor Agents — the construction-side discipline this maintenance discipline sits on top of.
- The Verification Frontier — “check the proof” is the operator-side instance of the verification problem.
- Agent SDK — How the Agent Loop Works — the harness primitives (tools, permissions, memory, stop conditions) made concrete in the SDK.
- Agents & Agentic Systems
- Build for the Next Model — a proposed answer to this article’s open question: compensating harness code expires, connecting code is what you maintain.
Try It
- Run the 5-point audit on one agent you already operate. Write the answers down. The “check its reach” and “check the proof” steps usually surface the most surprising drift.
- Do a tool-pruning pass. List every tool your agent can call, then ask of each: would removing it make the agent worse? Delete the ones where the answer is no, and re-measure.
- Write the “when the model changes” plan. For your deepest custom harness, note which guardrails exist only because the old model was clumsy — those are the first candidates to relax when you upgrade.
Open Questions
- The ~80% tool-deletion figure for Vercel’s agent is stated in the talk without a primary source link — worth corroborating against a Vercel writeup before treating the exact number as established.
- Jones frames the harness as the durable asset; Canva frames it as disposable. The reconciliation (which parts of a harness are durable vs throwaway) is an open synthesis candidate.