Production essays. Agents that finish the job.
Deep notes from Sunrise Gen AI on harnesses, inspectable memory, and the stops that keep a loop honest. We build production GenAI for operations, knowledge, and data. For the work itself, start at sunrisegenai.com.
Two speeds of a token: prefill, decode, and the latency you actually ship
Prefill is a compute burst that sets time to first token. Decode is a memory-bandwidth walk that sets inter-token latency. Shipping GenAI means budgeting both, plus the cache and the batcher behind them.
Security is a release gate, not a last look
When agents write most of the code, pre-ship security cannot be a vibe check the night before launch. Secrets, authorization, data rules, abuse limits, supply chain, and logs become numbered gates the harness runs on every release.
Brand is a skill file, not a vibe
Coding agents invent a new look every run unless taste lives as a loadable skill. Tokens, type, spacing, forbidden patterns, and a verify gate so the studio stays the same.
The agent file is the contract
Most production agents are Markdown contracts, not framework code. Role, instructions, constraints, tool refs. Describing the agent is the product surface. Description without a harness is still a costume.
Cache hit rate is a harness quality metric
Prompt stability, tool-trace shape, and cache reuse are how you score a harness. An engineered layout earns hits. A vibe layout burns the floor.
Ship know-how as skills, not as another prompt
A reusable skill is a named contract the harness can load. Role, constraints, tools, and a recipe that survives the next project. Slop skills are prompts in a costume. Production skills are how the loop stops relearning the same job.
Coding agents need numeric gates, not vibes
Verify for coding agents is a stack of measurable contracts in the harness. Pick the gates that fit the job. Fail closed. Retry or escalate. Never ship on confidence theater.
Slim context packs beat written rules
Long policy in the prompt is not control. Slim packs plus hard stops — allowlists, budgets, stop sequences, refuse-when-unsourced — are control. Evaluation is product surface.
Give the agent a filesystem it can inspect
Opaque vectors and session blobs make memory a black box. Hierarchical files, portable over MCP, are how a person opens the last write before the next action fires.
An agent is a model plus a harness
A model generates tokens. An agent finishes work. The difference is the loop, the tools, the permissions, and a test gate in front of done — the harness is the product.