Byline: Alchemic Technology. The summaries below are descriptions of what those posts say. They are not new case studies, and they are not measurements from a deployment. Where a summary depends on a paper or a specification, the primary source is linked so you can check the sentence.

On 6 November 2026 the posts in this guide were given source links for the documents they already named. Several of them had named those documents and had not linked them. This page is the map. The About page says how the blog is written, including the limit that many older posts still have no external link.

1. Permissions and authorization

Start here if the question is what an agent may cause, as distinct from what it can see.

Your tool output is not an authorization

A tool result is easy to treat as neutral data. The post’s point is that a result can stop being data and start reading like the next command: a wiki page that says to run a migration, a ticket that says to reveal an account, a log that says to disable a failing check. It separates action induction, the model inferring a possible next step, from execution authorization, the runtime deciding that step may happen.

Primary source: Zhao and coauthors, “When Tool Outputs Become Commands,” arXiv:2608.27146. The abstract argues for separating those two roles. The post also uses the MCP authorization specification for a narrower claim: permission to call a restricted server is not permission to perform every action the payload suggests.

The post treats an approve-or-reject dialog as incomplete once an agent is long-running and can resume. It argues for a record of who granted which action, what they were shown, which tool asked, when the grant expires, and what ran after the run resumed. The fields it lists include the proposed action, the decision, the scope, and a pointer to the paused run.

Primary source: OpenAI Agents SDK, Human-in-the-loop. That page says tools can declare that they need approval, that pending approvals surface as interruptions, and that run state can be serialized and resumed after a decision. The post also cites MCP elicitation (2025-06-18), which says servers request structured data through the client, must not use elicitation to request sensitive information, and that applications should offer decline and cancel.

Revocation has to fence old authority

The post’s distinction is between a kill switch and a revocation fence. Canceling a run, exiting a process, or revoking a credential can leave work that was already accepted: queued jobs, callbacks, reservations, and provider-side operations. The post says the receipt should name what was inspected and should leave missing or conflicting evidence marked indeterminate.

Primary source: “Authorization Revocation for Long-Running AI Agents,” arXiv:2609.21284. The abstract defines root-scoped authorization quiescence and says missing or conflicting evidence remains indeterminate. The post uses that paper’s language. It is a summary of the abstract and a design argument, not a reproduction of the proofs.

2. Failure handling

These three are about a run that has already started. Quarantine narrows what it may do. Cancellation says what “stop” means after a side effect. Compensation is the plan for a step that partly succeeded.

Quarantine is a state, not a better error message

The post argues that the dangerous moment is the step after a suspicious success: the tool returned HTTP 200, the schema validated, and the workflow’s bias is to continue. Quarantine, in the post’s usage, is a reversible, logged, capability-reduced state. High-impact actions stop. The run and its evidence stay.

The post includes an illustrative refund walkthrough. The order and the emails are made up. It is labeled as an illustration, not a customer incident. Primary source for the authorization claim: MCP authorization, 2025-06-18. The post also links the OpenAI guardrails docs, the OpenTelemetry GenAI conventions, the Claude Code hooks reference, and NIST’s four-function description, and it says which part of the argument each document actually supports.

Cancellation is a contract, not a button

The post says a stop control ends an interface state, and a cancellation contract has to say which phase the run was in, whether a side effect had already committed, and which receipt supports that status. “Unknown until verified” is one of the statuses it wants the product to be willing to show.

Primary source: MCP cancellation utility, 2025-06-18. Either side can notify that an in-progress request should be cancelled. The post treats that notification as a signal, and it says a signal is not proof that a payment, a CRM write, or a vendor API stopped.

Partial success needs a compensating action

A retry assumes the first attempt did nothing. The post’s case is the attempt that failed from the agent’s point of view and still changed a system of record: a note appended, an order reserved, an email sent. It asks for a detector other than the model’s final sentence, an idempotent repair, and a receipt that says what remains unresolved.

Primary source: Azure Architecture Center, Compensating Transaction pattern. That page describes undoing work when a step fails in an eventually consistent operation, recording progress so recovery can resume, and designing steps as idempotent commands. The post applies that pattern to agent tools. It is not an Azure implementation guide.

3. Audit and traces

A trace needs a retention budget

The post grants that the first question after a failure is “what did the agent see?” and then argues that keeping every prompt and tool payload forever creates a second store of customer data. It separates a durable metadata ledger from short-lived debugging payloads, an incident hold with an owner and an expiry, and fields that should be redacted before export.

Primary source: OpenAI Agents SDK, Tracing, which describes records of generations, tool calls, handoffs, guardrails, and custom events. The post also points at OpenTelemetry’s GenAI semantic conventions for shared span and event names. Retention periods are the post’s proposal. They are not a number in either of those documents.

How to use this page

Read the post, then open the primary source next to the paragraph that names it. If the post’s sentence is stronger than the source, the source wins. The source sections added on 6 November 2026 are there for that check.

This guide is one cluster. The blog has many other posts, including single-paper summaries and a few posts that state original measurements. Those are outside this page. The About page describes which revised posts link their sources, and it points at the audit script for the posts that still have none.

← Back to the blog