AI Context
Sense, Decide, Act, Close: Anatomy of an AI Agent Workflow
An AI that gives you advice is useful. An AI that closes the loop — reads context, reasons, takes an action, and logs the result — is how work actually gets done.
2026-06-26 · 8 min read
What makes an AI agent workflow different from a chat session
Chat AI is fundamentally request-response: you ask, it answers, nothing changes in your systems. An AI agent workflow is different in kind, not just degree. The model reads live context from your tools, reasons over it, writes actions back to those tools, and the results feed into the next loop. The AI is not advising you about your work — it is doing a defined piece of it.
This distinction matters because most of the overhead in knowledge work is not decision-making. It is the connective tissue around decisions: gathering the context to decide, communicating the decision, updating the systems that track it, following up when something slips. An agent workflow handles that connective tissue. You supply the judgment; the loop handles the logistics.
The Sense-Decide-Act-Close loop is Lodestar's model for how AI agent workflows should be structured. Each stage has a clear responsibility, and together they form a closed cycle that actually finishes work rather than adding more advice to your stack.
Sense: reading context before reasoning
Sense is the stage where the AI reads the context it needs to reason well. In Lodestar, this means calling MCP read tools: pulling your dashboard, reading action items with their evidence links, fetching email thread content, checking calendar events, querying the knowledge graph for relevant docs or prior decisions.
The quality of the Sense stage determines the quality of everything that follows. A model reasoning over a full email thread, the history of who owes what, and the sentiment trend for this contact will reach better conclusions than one reasoning over your summary of the situation. The context layer is what makes Sense meaningful rather than superficial.
Lodestar's enrichment pipeline runs before the AI ever touches the data. By the time you ask a question, your emails have already been classified and linked to projects, action items have been extracted with their source quotes, and sentiment has been calculated per contact. The model is not doing raw text analysis — it is reading structured, enriched context. That compresses the Sense stage from minutes to seconds.
Decide: where the model earns its keep
Decide is the stage most people picture when they think of AI: the model reasons over context and determines what to do. In a well-constructed agent workflow, Decide is where the frontier model's reasoning capability is applied to a problem that is actually worth its computational cost — not reconstructing context you already had, but making a judgment over it.
The Decide stage for a morning planning workflow might look like: "Given these seven overdue items, these three incoming threads, and this calendar, what are the three highest-leverage things to do before noon?" The model weighs urgency, relationship risk, and dependencies — the kind of multi-factor judgment that does not reduce to a rule.
For a client follow-up workflow, Decide might be: "This sentiment trend has dropped two weeks running, the last three emails went unanswered, and the renewal is in six weeks. Draft a recovery email and propose two follow-up tasks." That reasoning requires context (Sense), judgment (Decide), and the ability to act on the conclusion — which is the next stage.
Act: writing results back through MCP
Act is where the loop becomes real. The model calls MCP write tools: action_items_create to log a new commitment, emails_reply_link to open a draft in your mail client, milestones_partial_update to record that a phase completed, timeline_events_create to log what happened and when.
The design principle in Lodestar's Act layer is that consequential actions surface for human review. Email drafts are exactly that — drafts. You see the proposed reply, decide it is right, and click send. The AI is doing the drafting work; you retain the send decision. Task creation and milestone updates happen immediately, since they are reversible and scoped to your workspace, but even those are logged to the timeline so you can see what the AI did.
This is not a safety compromise — it is the correct design for a professional workflow. An AI that cannot be overridden is not useful in a context where judgment calls depend on information the model cannot see: relationship dynamics, political nuance, information shared verbally. The Act layer executes the mechanical work; the human judgment gate stays in place for the communication.
Close: logging, notifying, tracking until done
Close is the stage that prevents work from evaporating. After the model acts, Lodestar logs the event to the timeline (what happened, who was involved, what was the outcome), updates the relevant project and milestone state, and creates or closes action items so the tracking picture reflects reality.
The Close stage also handles follow-through. If a follow-up task was created in the Act stage, Lodestar tracks it through to completion — surfacing it in the daily plan when it is due, flagging it in the needs-attention feed if it slips, and automatically nudging until it is closed or deliberately deferred.
This is why the loop is called Sense-Decide-Act-Close rather than Sense-Decide-Act. Without the Close stage, you have a model that takes actions but leaves no trail. With it, every AI action is auditable, every commitment is tracked, and the context layer grows richer with each completed loop — making the next Sense stage more accurate.
A complete loop: builder ships a feature
Walk through a concrete AI agent workflow for a solo developer using Claude Code with Lodestar. Start of session: Claude calls dashboard_today and action_items_list. Sense stage returns three items due today, one blocking the Jet Rebellion feature. Claude surfaces the plan with priorities.
The developer says: "Implement item 42." Claude calls action_items_retrieve(42), then knowledge_search and files_content to pull the spec and relevant code files. Sense is complete. The model reasons over the spec, the existing code, and the blocking issue. Decide produces an implementation plan. Claude Code writes the code, runs tests, deploys.
Act: Claude calls action_items_partial_update(42, 'done') and timeline_events_create to log the completion. It calls emails_reply_link to open a draft notifying the stakeholders who were CC'd on the original spec thread. The developer reviews the draft and clicks send. Close: the action item is done, the milestone is updated, the timeline records the event, and the stakeholders are notified — all without leaving the editor.
A complete loop: consultant recovers a client relationship
A fractional executive starts the morning with a question to Claude: "Anything on fire?" Sense: Claude calls dashboard_needs_attention and dashboard_sentiment. Returns one client with a sentiment drop of −0.3 over the last two weeks, two unresolved threads going on ten days without response.
Decide: Claude reads the email threads (emails_content), the org chart (knowledge_org_charts), and the open action items for this client (action_items_list). It identifies what went sideways: a deliverable missed, a question left unanswered, a contact who has gone quiet. It drafts a recovery approach.
Act: emails_reply_link opens a draft to the right contact addressing the situation directly. action_items_create logs two follow-up tasks with owners and due dates. Close: timeline_events_create records the decision to re-engage. The daily plan will surface both follow-up tasks until they are marked done. What would have taken twenty minutes of manual reconstruction — reading threads, checking notes, writing a reply from scratch — runs in one conversation.
Why the loop matters more than any individual AI capability
The power of the Sense-Decide-Act-Close model is not any single stage — it is the loop. A model that can reason well but cannot read structured context gives generic advice. A model that can read context but cannot write back to your systems gives better advice, still manual. A model that reads, reasons, acts, and closes produces finished work: committed, tracked, and followed through.
For independent operators and technical builders, the compounding effect matters most. Every completed loop enriches the context layer. A client relationship with fifty logged interactions, closed commitments, and a tracked sentiment trend is a fundamentally different input to the next Sense stage than a blank slate. The AI gets more useful as the loop runs — not because the model improves, but because the memory it reads from grows.
This is what it means for AI to actually do work rather than assist with it. The architecture is available today through the Lodestar MCP server and any compatible AI client. The loop is already defined. The context layer is already being built. What you bring is the work.
The takeaway
The Sense-Decide-Act-Close loop is the framework that turns a capable model into one that actually finishes work. Early access to Lodestar gives you the context layer and MCP server to run this loop from day one — with a product team available to work through your specific workflows during onboarding.
Context explained
Why context changes everything about AI — and how Lodestar builds the memory layer your models are missing.
Read Context Explained