From read-only to write actions: release an AI agent's changes in stages
Release AI agent write actions in stages: proposals, approved execution with receipts, narrow automatic actions, server-side limits and evidence to advance.
Field notes
Production AI patterns we ship every week. Stuck pilots, MCP hardening, eval suites, agent orchestration, cost control. Written by Alex Cinovoj, Founder & CTO, TechTide AI · CTO & Co-Founder, CPG Playbook. No vendor pitches, no recycled blog posts.
71 of 71 field notes · 2026 archive
Release AI agent write actions in stages: proposals, approved execution with receipts, narrow automatic actions, server-side limits and evidence to advance.
What to require at the end of an automation project: inventory, owned credentials, written rules, health signals, a rehearsed runbook and test evidence.
A proposed RMA process: order matching, quoted return terms, safe evidence uploads, separated complaints and return decisions made by authorized staff.
A proposed handover workflow using a COBie-style structure: data needs per asset type, schedule matching, sourced attributes and owner-recorded acceptance.
A proposed workflow for lease options and notice windows: dates stored as sourced computations, amendment checks, notice requirements and escalating reminders.
Confidentiality questions for a law firm AI pilot: training, retention, access, ethical walls, output checks, data rules and exit, grounded in ABA guidance.
How to move an AI workflow from full human review to sampling: written criteria, reviewer agreement, per-risk groups, staged reductions and reset triggers.
A proposed business credit workflow: completeness checks, sourced financial figures, organized references, analyst-owned decisions and approved notices.
A proposed professional services workflow: transcript-quoted action items, unclear owners left unclear, scope flags and nothing sent to a client unreviewed.
A CRM migration plan built on reconciliation: scoped objects, field maps by meaning, source IDs, people-led duplicate merges, rehearsal loads and sample checks.
A proposed estoppel workflow for sales and refinancings: a sourced lease baseline, field-by-field comparison of returned certificates and recorded decisions.
A proposed engagement letter workflow: approved templates, facts from confirmed intake, visible gaps, no generated fee terms and a recorded lawyer approval.
Questions to ask before hiring a workflow or AI automation consultant: account ownership, scope, testing, failure handling, AI decisions, pricing and handoff.
A proposed submittal workflow: confirmed spec checklists, sourced completeness findings, original files intact and engineering judgment kept with reviewers.
Move an n8n integration from prototype to production: business-owned credentials, separate environments, authenticated webhooks, error workflows and owners.
A proposed workflow for Columbus benchmarking: meter inventory, sourced bill data, consistency checks and a review package before Portfolio Manager submission.
A proposed docketing workflow: AI extracts trigger events and periods with source quotes, code computes dates, and a docketing reviewer confirms each entry.
Plan a CRM integration field by field: one owning system per value, sync direction, stable record IDs, deletion rules and visible sync failures.
Copilot optimizes for keystrokes, Claude Code optimizes for tasks. Which to choose, where each fits, and how they differ on MCP tools, review, cost and governance.
How to measure an AI workflow by cost per accepted outcome, with rework, retries and operating effort included and compared with the previous process.
A proposed delivery-packet workflow that checks issued files against reviewed revisions, approvals and the transmittal so the package can be reconstructed.
A proposed staffing workflow that suggests feasible assignments from current availability with visible assumptions, leaving the decision to a delivery manager.
A proposed baseline-first measurement method: one workflow, consistent definitions, and operating costs reported alongside the measured benefit.
A proposed sales handoff workflow that gives delivery staff the accepted commitment, its source document and the person who approved any exception.
AI-generated prototypes are the cheapest product research available. The failure is promoting one to production without tests, types, auth, or an owner. The handoff checklist we use.
A proposed portfolio reporting workflow that confirms which properties, periods and source files are present before totals or narrative are produced.
A forecast is not an individual diagnosis. Review ownership, costs, tests, expectations, and use-case fit against the project's evidence.
A proposed internal legal knowledge search that starts with access tests: partial access, lost access, team changes and cached answers.
MCP is vertical, model to tools. A2A is horizontal, agent to agent. Choose based on whether you need tool access or coordination between independently operated agents. The decision rule, with failure modes.
An illustrative planning sequence: review the pilot, address its foundations, and decide whether the acceptance evidence supports release.
How to write AI pilot acceptance criteria before the demo: observable behavior, the evidence that checks it, and what happens when the workflow cannot finish.
Six months after the first AI feature ships, the debt shows up as prompt sprawl, untyped tools, no eval baseline, and nobody able to say what changed. How we pay it down.
A proposed field-report workflow that turns photos, notes and forms into a reviewable draft, keeping who observed what and leaving conclusions to engineers.
A proposed renewal packet that separates delivered, accepted, client-reported and measured results from open issues, without a generated success story.
Model swaps stopped moving the needle. The gains now come from the harness around the model: durable state, checkpoints, retry policy, budget ceilings, and clean recovery.
A proposed order-change workflow that establishes the current accepted order, checks the affected systems and collects approval before applying the change.
The agent does not need to be right every time. It needs to know when it is out of depth. Four escalation triggers, plus the handoff payload that makes a human fast.
A proposed lease abstraction workflow where every extracted field carries its source passage, document version and open interpretation questions.
Leaderboards measure someone else's problem. Build a 50-task set from your own traffic, score it blind, price it per task, then keep the swap cheap.
A proposed matter-closeout workflow that reconciles the checklist across systems and reviews access before anything changes, under the firm's own rules.
How to scope a first Claude MCP integration around read-only questions: narrow tools, the user's own access, source links, freshness and an evaluation set.
Enterprise AI governance frameworks are unusable at 200 people. Here is the one-page version we install: data boundaries, approved models, logs, review gates, named owner.
A proposed review-comment workflow that suggests duplicate groups while keeping every original comment, its revision and an authorized closure decision.
Every long-context release restarts the RAG obituary. Retrieval still wins on cost, freshness, permissions, and citation. The five fixes that rescue a weak pipeline.
A proposed client reporting workflow that drafts status reports from bounded project evidence and leaves missing facts visible for the delivery lead.
A proposed accounts payable workflow that matches invoices against purchasing records and routes exceptions to people before any payment is authorized.
Filtering instructions out of untrusted text is a losing game. The defenses that hold are boring: least-privilege tools, typed arguments, allowlists, and approval on writes.
A proposed vendor document workflow with separate requested, received, under review, accepted, rejected and superseded states, so receipt is never approval.
Swarms demo well and page you at 2am. Three orchestration patterns survive real traffic, and the choice is decided by failure isolation, not elegance.
A proposed review queue for legal billing narratives that flags missing context and asks precise questions, without inventing work or changing time.
How to write n8n recovery steps operators can follow: alerts that point to a decision, failure classes, deliberate retries, preserved evidence and rehearsal.
A proposed transmittal workflow that reconciles documents, revisions, approvals and recipients into a reviewable packet before distribution.
Debugging an agent from a chat transcript is guesswork. Log the full run as a trace, keep cost and latency on the same timeline, and alert on tool-level failures.
A proposed scope-change workflow that compares client requests with the current agreed scope and prepares the decision for the delivery owner to approve.
A proposed purchasing worksheet that keeps quote differences, source locations and clarification questions visible instead of ranking incomplete quotes.
Teams argue about whether the agent should be autonomous. Wrong unit. Set autonomy per action, using reversibility and blast radius, and earn each step up with evidence.
A proposed maintenance routing workflow that keeps the tenant's original words, applies the property's escalation rules and shows uncertain cases to a person.
A proposed legal document request workflow: one record per requested item, original files kept, missing items visible, and counsel deciding sufficiency.
Nobody plans an AI budget overrun. They just never set a ceiling. Per-task budgets, tiered routing, context hygiene, caching, and one dashboard with a named owner.
Design a usable review queue with named approvers, exact action previews, expiry, escalation, and execution receipts.
Fourteen nodes in a state graph to answer refund questions. Size the stack to the problem: one loop, typed tools, a log, and only then a framework.
A practical data-flow review for client-service teams adopting AI: access, retrieval, logging, retention, and deletion.
Every MCP server ships two products at once: the tools and the attack surface. Fourteen checks on auth, scoping, tenancy, argument validation, and audit trails.
A proposed workflow for collecting change evidence, checking baselines, routing impact reviews, and preserving engineering approval boundaries.
No eval suite means every change is a vibe. Five days: harvest real tasks, write assertions, set the pass bar, wire CI, then gate deploys on it.
A proposed onboarding workflow connecting approved scope, client inputs, access requests, and delivery readiness without inventing commitments.
The prompt is the smallest part. Production quality comes from context assembly: retrieval scope, ordering, compaction, tool results, and a hard token budget.
A proposed RFI intake workflow that preserves drawing references, routes questions, and keeps contractual decisions with authorized reviewers.
A proposed workflow for transaction document intake, version tracking, missing-item review, and controlled reminders.
MIT says 95%, Gartner says 40%, vendor decks say 88%. They measure different things. What each stat counts, and the five gaps that separate shipped from shelved.
A proposed legal intake workflow with source-linked facts, conflict-review routing, confidential data boundaries, and human approval.
Buying guides on alexcinovoj.com and engineering notes on alexcinovoj.dev, published alongside these field notes.
Companion reading
How to read an implementation quote: process clarity, data preparation, integrations, review and recovery, and operating ownership. A proposed evaluation method, not a price list.
Published October 29, 2026 · January 2026 collection
A decision guide for choosing between software you already own, an extension and a custom integration, with exit paths and a bounded trial before expanding. The examples are hypothetical.
Published October 29, 2026 · February 2026 collection
A procurement method that asks every supplier the same questions about scope, working evidence, acceptance criteria, security boundaries and handoff. It does not rank agencies.
Published October 29, 2026 · February 2026 collection
A walkthrough of ClawKeeper's deterministic agent policy, tenant checks and payment approval gate, based on seven synthetic policy tests run at one inspected revision, with its known trust boundaries.
Published October 29, 2026 · January–March 2026 collection
A source-code reading of CipherClaw's trace graph, root-cause heuristics and breakpoints at one inspected revision, with its implementation limits. The review did not run the test suite.
Published October 29, 2026 · January–March 2026 collection
Service areas
Columbus metro in person. Remote across the US, Canada and the EU.
© 2026 TechTide AI LLC. Columbus, Ohio.