NateWiki — Open Brain corpus
Home · The Doctrine · All topics · Search
priority 1 62 posts

Agent Architecture & Orchestration

The playbook

Nate's core, unchanging claim across two years of writing: an agent is nothing but "LLM + tools + guidance," and every success or failure traces back to the harness wrapped around the model — loop structure, memory strategy, tool registry, action constraints, execution environment, accountability layer — not to model quality. Wells Fargo ran 245 million agent interactions cleanly while MD Anderson lost $62M on IBM Watson and McDonald's drive-thru went viral for the wrong reasons, using comparable-era models. The difference was architecture and scope discipline, not intelligence.

That harness thesis plays out through three rough eras. 2024 into mid-2025 is protocol and ecosystem-mapping: Claude ships MCP (Nov 2024) and Nate treats it as the first real standard for tool access; by March 2025 he's mapping OpenAI's closed Agents SDK against Claude+MCP against LangGraph as three distinct lock-in/control tradeoffs, and dismissing early Operator/computer-use as "pretty bad" GUI-clicking months before that's confirmed in a head-to-head test (Comet beat Operator 8-0 on real workday tasks in July 2025 — OAuth/API-native design beat vision-based screenshot navigation regardless of underlying model). Mid-to-late 2025 is architecture and context-engineering: the June 2025 "Definitive Guide" locks in a three-tier memory model (working/episodic/long-term) and a single-agent-first default (multi-agent costs 3-10x more in tokens and time, justified only by genuine parallel-expertise needs); the September 2025 MCP piece gives hard numbers for where MCP breaks (300-800ms latency, 5,000-50,000 ops/sec ceiling, 3.25x-236x token blowup on injected context, a real 34-day Asana cross-tenant breach from baked-in token passthrough) — MCP belongs in background/intelligence work, never a hot transactional path. By December 2025, after synthesizing Google/Anthropic/Manus papers, he corrects his own prior belief that bigger context windows would fix agent degradation: it's architectural (attention dilutes as tokens accumulate), so the fix is computing context fresh each call from structured state, not pinning everything into a mega-window.

2026 is the governance and commerce era, catalyzed by the OpenClaw/Moltbot saga (a viral open-source always-on agent runtime that went from "chatbot wrapper" to genuine durable runtime between February and May, while racking up a real security record — 21,639 exposed instances found on the open internet). That scare reshapes everything downstream: Anthropic's native /loop scheduling plus owned memory ("Open Brain") is explicitly pitched in March as "OpenClaw you control"; the April leak of Claude Code's source (512K lines, 1,902 files) proves the LLM call is only ~20% of a working agent, the other 80% being unglamorous plumbing — session persistence, permission tiers, idempotency, token-budget pre-checks, structured logging; May's "Judge Layer" piece (following Lindy's public postmortem) formalizes that actor and judge must be structurally separate roles, with a four-outcome verdict (Allow/Block/Revise/Escalate) gated by action risk tier. By June, agent ownership (a named human accountable per agent) and standardized handoff contracts (task record: requester, outcome, sources, acceptance criteria, boundaries, blocker rule, receipts) become the operating unit, and Nate reframes the whole two-year arc as "loops" — bounded recurring responsibilities with a trigger, memory, safe actions, and a judgment boundary — arguing this is the same tools/memory/goals/harness vocabulary from his technical writing, just brought down to weekly-task scale.

Running under all three eras: agent-as-adversary security thinking (assume the agent can be compromised or manipulated, sandbox and scope before granting real permissions); a repeated numeric argument that reliability compounds against you (95% per-step accuracy means under 10% success over a 50-step task, which is why verification loops matter more than smarter models); and a consistent "graduated autonomy" stance against full-autonomy hype — agents act freely at low risk, seek approval at moderate risk, prepare human-reviewable evidence at high risk or irreversibility, never all-or-nothing.

Key moves

The posts

Related