Prior authorization is an evidence workflow
A synthetic workbench follows a missing record through request, reassessment and human review, with precise limits on what its agent can decide.
4 min read
agents.md · 23 posts
A synthetic workbench follows a missing record through request, reassessment and human review, with precise limits on what its agent can decide.
4 min read
A practical release evaluation that checks outcomes, authority, repeated runs, latency, and cost. Includes a small harness that exposes a misleading improvement.
8 min read
A field-service workbench that checks parts, people and access before a visit, and withdraws approval when its supporting evidence changes.
4 min read
What a tool should remember when a request succeeds but its response disappears. A runnable SQLite example covers duplicate delivery, conflicting intent, and concurrent retries.
8 min read
An invoice-dispute workbench that connects evidence, exact human approval and verified credits, including recovery when a response is lost.
4 min read
A changed delivery date leaves a buyer with stock, cost and approval decisions. This experiment turns supplier terms into a traceable recovery plan.
5 min read
A practical way to find where an agent belongs in a business process, using a workflow map, explicit decision boundaries, and a runnable support example.
10 min read
A merged patch does not tell a security team what is running. This experiment follows a vulnerability through deployment evidence, human review and rollback.
4 min read
Naming a phase is not the same as discovering one. "Graph engineering" is the current label for wiring agents into explicit execution graphs. The essay…
1 min read
A new label is going around for building agentic systems as explicit execution graphs. The label is barely real. The shift underneath it is, and the evidence for and against it is more interesting than either side admits.
12 min read
The planner, coder, judge split is a good harness: an expensive model plans, a cheap one writes, the expensive one judges, with the budget spent where judgment lives. But an LLM judge is a filter, not an alibi; it catches the cheap model's misses, not whether the plan solved your problem or whether the edge cases the models quietly agreed on match the ones your users will find. The one step that does not delegate is the last: somebody opens the diff and reads it.
1 min read
Agentic AI is in its demo phase, where a chain of prompts looks magical in a notebook but falls apart in production. The discipline that separates the two is knowing where reasoning belongs to the model and where reliability belongs to ordinary code.
11 min read
The agentic future is now mainstream consensus, and tokens have become the unit of account for the intelligence economy. But tokens are a unit of labor, not money; the real money between agents is stablecoins; and the most actionable move is building software agents can consume.
13 min read
A provider's SDK earns its keep on six concerns: auth, request shape, transport, retries, streaming reassembly, and response parsing. Everything above that, the agent loop, planning, routing, memory, parallelism, guardrails, and workflow state, is application architecture and is better written in pure Python.
16 min read
The lunch-table conversation about AI agents and the production-engineering conversation about AI agents are both right. They are talking about three different things and calling all of them "agent."
10 min read
An agent with no memory repeats itself. An agent with bad memory contradicts itself. Getting memory right means splitting it into kinds: short-term for the session, long-term for durable facts, procedural for how-to, episodic for what-happened.
8 min read
LATS brings Monte Carlo Tree Search, the algorithm behind AlphaGo, to language agents. Reasoning, acting, and planning become a single tree that the agent navigates by balancing exploration against exploitation. Accuracy climbs, cost climbs with it.
8 min read
When an agent fails a task, a separate critique call can write down what went wrong in words, store the critique as a memory, and let the next attempt condition on it. No gradient updates. A model that learns from text about its own past.
8 min read
Anthropic reports spending more time tuning tool descriptions than tuning prompts for a SWE-bench agent. Tool design is where agentic systems actually succeed or fail, and it pays more compounding returns than almost any other investment.
7 min read
ReAct interleaves planning with acting at every step. Plan-and-execute produces a complete plan first, then executes it. The separation trades adaptivity for token efficiency and is often the better choice when tasks have clear structure.
7 min read
The 2022 ReAct paper parsed thought-action-observation from free text. Production agents today use native tool-calling, a step budget, input and output guardrails, and explicit handoff handling. The simple loop earns its hardening.
7 min read
The foundational agent pattern from Yao et al. interleaves a short reasoning trace with a single tool call, then folds the observation back into the next step. Simple in shape, strong on grounded reasoning benchmarks.
7 min read
The term agent has become overloaded, and the overloading is expensive. A system in which the language model controls its own flow costs more to run and more to debug. The default should be a workflow.
7 min read