https://sugeerth.github.io/ Essays on LLM systems, agents, recommendation, and evaluation — written from what actually broke in production. en-us Sun, 09 Aug 2026 23:45:00 +0000 https://sugeerth.github.io/blog/last-click/ https://sugeerth.github.io/blog/last-click/ An agent fixed a real open-source bug in 40 minutes. Shipping it took 5 hours, 4 permission walls, and exactly one human click — the full trace, rendered live, and the design lessons. Sun, 09 Aug 2026 23:45:00 +0000 https://sugeerth.github.io/blog/harness-house/ https://sugeerth.github.io/blog/harness-house/ Swap the model five times — same product. A razor for what survives (model-shape vs world-shape), the swap protocol, and a live 36-month simulator. Thu, 06 Aug 2026 04:50:02 +0000 https://sugeerth.github.io/blog/generative-ui/ https://sugeerth.github.io/blog/generative-ui/ Generative UI, with a fully working playground: request → intent → editable spec → live rendered interface. No server, no canned screenshots. Tue, 04 Aug 2026 05:28:06 +0000 https://sugeerth.github.io/blog/agent-hippocampus/ https://sugeerth.github.io/blog/agent-hippocampus/ The four-memory architecture agents are missing — episodic, semantic, procedural, working — and the nightly consolidation loop that turns experience into skill. Tue, 04 Aug 2026 05:12:41 +0000 https://sugeerth.github.io/blog/agent-loop-infra/ https://sugeerth.github.io/blog/agent-loop-infra/ The loop as an OS: memory management, resource budgets, watchdogs, checkpoints, and evals for runs too long to rerun — with a live run console. Sun, 02 Aug 2026 19:20:21 +0000 https://sugeerth.github.io/blog/domain-agents/ https://sugeerth.github.io/blog/domain-agents/ The full playbook: data flywheel, training ladder, K verified paths, golden evals, ratcheted deployment — and why the moat is the eval set, not the model. Sat, 01 Aug 2026 05:12:04 +0000 https://sugeerth.github.io/blog/autonomy-dial/ https://sugeerth.github.io/blog/autonomy-dial/ When should an agent ask vs. act? An expected-loss framework, a trust ratchet, and a live policy observatory. Wed, 29 Jul 2026 17:38:54 +0000 https://sugeerth.github.io/blog/seeing-agents/ https://sugeerth.github.io/blog/seeing-agents/ Visualization is the missing layer of the agent stack — five layers where it plugs in, plus a live control room with pluggable modules. Wed, 29 Jul 2026 05:09:19 +0000 https://sugeerth.github.io/blog/agent-tool-design/ https://sugeerth.github.io/blog/agent-tool-design/ The seven sins of tool design for AI agents, and a live grader that scores your tool definitions as you type. Sun, 26 Jul 2026 23:44:53 +0000 https://sugeerth.github.io/blog/long-horizon-eval/ https://sugeerth.github.io/blog/long-horizon-eval/ Final-answer evals can't tell long-horizon agents apart. The discriminating signal lives in the trace — a metric framework, behavioral clustering, and a live D3 lab. Sat, 25 Jul 2026 03:54:31 +0000 https://sugeerth.github.io/blog/agent-memory/ https://sugeerth.github.io/blog/agent-memory/ The models crossed the endurance threshold. The memory engineering around them didn't. A field guide to agent memory for long-running tasks — via Anthropic's harness and Hermes. Sun, 12 Jul 2026 23:28:42 +0000 https://sugeerth.github.io/blog/goodhart-in-the-act/ https://sugeerth.github.io/blog/goodhart-in-the-act/ Watching a self-improving AI loop start gaming its own grader, live — and how fast the proxy-vs-truth gap lets you catch it. Fri, 10 Jul 2026 18:15:06 -0700 https://sugeerth.github.io/blog/explainable-ranking/ https://sugeerth.github.io/blog/explainable-ranking/ A four-stage funnel exposes four different surfaces — a cosine scalar, a feature vector, a permutation, a free-text rationale. This is the method-by-method guide to explaining each, and the one rule underneath all of them: plausible is not faithful. Mon, 22 Jun 2026 22:44:42 -0700 https://sugeerth.github.io/blog/two-tower-retrieval/ https://sugeerth.github.io/blog/two-tower-retrieval/ The dual-encoder didn't win because it was the most accurate model. It won because its one sacrifice — no interaction until the final dot product — is exactly the property a billion-vector index needs to exist. The full story of that bargain. Mon, 22 Jun 2026 16:55:46 -0700 https://sugeerth.github.io/blog/agentic-ai-verification-gap/ https://sugeerth.github.io/blog/agentic-ai-verification-gap/ Across software, medicine, law, finance, science and more, AI agents hit the same wall — and the bottleneck was never generation. It's trust. Mon, 22 Jun 2026 02:38:25 -0700 https://sugeerth.github.io/blog/ai-agents-2026/ https://sugeerth.github.io/blog/ai-agents-2026/ Capability is soaring while deployment stalls. Why reliability — not intelligence — is the wall for AI agents in 2026. Mon, 22 Jun 2026 02:38:25 -0700 https://sugeerth.github.io/blog/multi-agent-coordination-tax/ https://sugeerth.github.io/blog/multi-agent-coordination-tax/ When many AI agents beat one — and the ~15x bill, failure taxonomy, and security worms that decide it. Mon, 22 Jun 2026 02:38:25 -0700 https://sugeerth.github.io/blog/agent-amnesia/ https://sugeerth.github.io/blog/agent-amnesia/ Most 'agent memory' is retrieval over a transcript. Real memory needs forgetting. A live simulation of why hoarding everything makes agents worse. Sun, 14 Jun 2026 21:52:26 -0700 https://sugeerth.github.io/blog/agent-no-undo/ https://sugeerth.github.io/blog/agent-no-undo/ Agent infra copied the model and skipped 50 years of database wisdom. A live simulation of why long agents fail silently — and how transactions fix it. Sun, 14 Jun 2026 21:52:26 -0700 https://sugeerth.github.io/blog/grading-ai-wrong-test/ https://sugeerth.github.io/blog/grading-ai-wrong-test/ A model that aces the demo can still fail in production. The gap is how you graded it. A field-grounded playbook for evaluating LLMs honestly. Fri, 12 Jun 2026 10:11:21 -0700 https://sugeerth.github.io/blog/agents-marathon/ https://sugeerth.github.io/blog/agents-marathon/ The hardest part of a long task isn't being smart. It's not falling apart on mile 22. A field-grounded playbook for building agents that finish. Mon, 01 Jun 2026 16:49:18 -0700