How to Maintain Context Across AI Workflows: The Decision Thread
How to Maintain Context Across AI Workflows: The Decision Thread
Say your team spent three weeks interviewing customers before a launch. The plan was to sell to enterprise operations teams and have pilots running within 90 days. The interviews showed you couldn't reach enterprise buyers that fast, while mid-market operations leaders were easy to reach and had the problem now. So you chose mid-market.
The next job is an onboarding experiment. You hand it to an AI agent with everything: the interview transcripts, the old launch plan, the research summary and the chat where the team argued it out. The agent comes back with a careful onboarding plan for enterprise accounts.
Nothing went missing. Every transcript was there, and so was the old plan.
That's what context loss usually looks like in AI work. The agent didn't forget a message. It never found out the team had dropped enterprise, or why.
Most teams respond by adding more: longer transcripts, more uploads, a bigger context window. We think the better answer is a short, current record of the decisions that govern the work, handed over first. We call it the Decision Thread.
Why more context doesn't keep the direction
More context changes what a model can see. Continuity depends on which part of it the model treats as current and in charge. In the launch example, the decision to go mid-market sits in one summary among dozens of transcripts, several of them enterprise interviews pointing the other way.
Researchers have measured a version of this. In "Lost in the Middle", published in Transactions of the Association for Computational Linguistics in 2024, Nelson Liu and colleagues tested language models on multi-document question answering and key-value retrieval. Performance was often highest when the relevant information sat at the beginning or end of the input and dropped significantly when it sat in the middle, even for models built for long contexts. The study tested the models of its time, but the gap it found still matters: putting information in front of a model doesn't make it govern the answer.
Memory across long conversations runs into the same gap. LongMemEval, presented at ICLR 2025, tests chat assistants with 500 questions across five abilities: information extraction, multi-session reasoning, temporal reasoning, knowledge updates and abstention. The authors report commercial chat assistants and long-context models "showing a 30% accuracy drop on memorizing information across sustained interactions."
Two of those abilities describe the launch problem. Moving from enterprise to mid-market is a knowledge update: the old plan is still in the record, and the system has to know it no longer applies. Abstention means saying when the answer isn't there. If the agent can't find which segment the team chose, it should ask, not pick one from the transcripts.
What the Decision Thread holds
A Decision Thread is a short note that travels with the work. It tells the next person or agent what the work is for, which sources are in charge, what's already decided and why, what's still open and who acts next. It links to the full evidence instead of replaying it.
It has six parts:
- Outcome: what result the work has to produce, for whom, and which decision or action it should make possible.
- Governing sources: which policies, data, research or instructions are in charge, and which versions.
- Fixed decisions: which choices are made and should be inherited rather than reopened.
- Rationale and change: why the big decisions were made, and what evidence changed the direction.
- Open questions: which assumptions, risks or gaps in the evidence are still unresolved.
- Next action and owner: what happens next, who owns it and where a person reviews or approves it.
Here is the thread for the onboarding experiment:
Outcome: Design a two-week onboarding experiment for five mid-market operations leaders.
Governing sources: approved positioning guide (product claims); July interview synthesis (customer language); approved segment criteria.
Fixed decision: The launch segment is mid-market operations leaders. Don't reopen it.
Rationale and change: Mid-market won on reachability and urgency, not total market size. The interviews showed we couldn't reach enterprise buyers inside the 90 days the old plan assumed.
Open question: Does the setup work in week one reduce activation?
Next action and owner: The growth lead reviews participant criteria before any outreach.
The transcripts are still there for anyone who needs them. The thread tells the agent which conclusion they led to.
Keep three kinds of context apart
The thread works best beside two other layers, not buried in them.
Durable knowledge outlasts any one project. For the launch team, that's the approved positioning guide and brand standards. They'll still apply to the next launch.
Project evidence belongs to one piece of work: this quarter's interviews, a customer dataset, campaign results, the files behind one deliverable. It can be in charge for this project without becoming permanent company truth. The enterprise interviews explain why the team changed course. They shouldn't steer the next launch just because they sit in the same folder.
Live decision state is the thread. For the onboarding experiment, it names the positioning guide as the authority on product claims and leaves the enterprise interviews out of the governing sources.
Mixing the layers causes two familiar failures. One is archive inflation: everything goes into one searchable pile until current policy, old drafts, brainstorms and contradicting analyses all look equally legitimate. That's how the enterprise interviews ended up carrying as much weight as the decision to drop enterprise. The other is prompt inheritance, where facts from an earlier project ride into a new run because someone reused the template or the conversation.
The NIST AI Risk Management Framework treats context as ongoing work. Its Map function begins with "Context is established and understood," and NIST says to keep applying it "as context, capabilities, risks, benefits, and potential impacts evolve." NIST is writing about AI risk, but the habit carries over: context isn't a setup step you finish once. When the interviews moved the segment, the context for all the work after them moved too.
Hand over a packet, not a transcript
When work moves to the next agent or person, send a short packet: the current thread, links to the evidence they may need, the output you expect and where review happens. For the onboarding experiment it might read:
Produce the participant criteria and outreach draft for the approved onboarding experiment. Mid-market operations leaders are the fixed segment; do not reopen selection. Use the July synthesis for customer language and the approved positioning guide for product claims. Flag gaps in the evidence instead of filling them in. The growth lead approves the criteria before any outreach.
Each sentence covers something the agent would otherwise guess: the next outcome, the decision that stays fixed, what each source is for, what to do when the evidence runs out, and where a person has to say yes before anything goes out.
It also leaves things out on purpose: the exploratory conversation that didn't change direction, the alternatives the team discarded and extra copies of evidence the agent can open by link.
Microsoft Research's multi-agent system, Magentic-One, splits this kind of state in two. Its Orchestrator keeps a Task Ledger of facts, educated guesses and the plan, and a Progress Ledger where it checks how the task is going and assigns the next subtask. When progress stalls, it updates the Task Ledger and makes a new plan. The launch team's thread carries both kinds: the settled segment, which shouldn't move, and the growth lead's review, which is what happens next.
OpenAI's practical guide to building agents describes a handoff in its Agents SDK as starting the new agent "while also transferring the latest conversation state." But hand the onboarding agent the latest conversation and it gets the enterprise argument along with the conclusion. The packet tells it which one counts.
Update the thread when the meaning changes
Don't update the thread after every message. Update it when the meaning of the work downstream changes. In practice, that's when:
- a person approves or rejects a recommendation;
- new evidence changes the plan;
- a source is replaced or stops being authoritative;
- a constraint, deadline, audience or risk changes;
- an open question gets answered;
- the work is about to leave the company;
- the work moves to a new specialist, tool or project.
Each update should answer three questions. What changed? Why did it change? What does the work after it now have to do differently?
"Updated plan" answers none of them. Compare: "Moved from enterprise to mid-market because the interviews showed we couldn't reach enterprise buyers inside the 90 days the plan assumed. The onboarding experiment now targets five mid-market operations leaders." That note changes what the next agent does.
Anthropic's engineering team ran into the same need in its work on long-running coding agents. "Each new session begins with no memory of what came before," they write, comparing it to engineers working in shifts. Their answer was a first session that sets up the environment, a progress file logging what agents have done, and descriptive git commits, so the next session can see where the work stands. The Decision Thread is the same kind of shift note, written for business decisions instead of code.
Say which source governs what
Finding relevant material isn't enough if the workflow can't tell a current authority from something that only looks similar. Say the launch team has three documents:
- an approved positioning guide from July;
- customer interviews from June;
- a product brainstorm from March.
All three are relevant to the outreach draft. They aren't equally in charge. The positioning guide governs product claims. The interviews supply customer language and evidence. The brainstorm explains some history, and it shouldn't quietly turn into a product promise because an agent found it while searching.
Without those roles written down, an agent can produce a confident blend of all three. Give each governing source a role, an owner, a version or effective date and, where it applies, what replaces it. When two sources conflict, the thread says which one wins and which one is supporting evidence.
Freshness works the same way. When the positioning guide changes, update it and find the decisions and active work that depend on it. A new version sitting somewhere in the knowledge base doesn't help until the thread points to it.
How to start a Decision Thread
Before a multi-step workflow begins, write six short lines:
Outcome:
Governing sources:
Fixed decisions:
Rationale and changes:
Open questions:
Next action, owner and review point:
Then keep it honest with four habits.
- Keep it short. Link to evidence instead of pasting whole documents.
- Record meaning, not motion. "Research completed" is status. "Research showed we had the wrong buyer" changes the next piece of work.
- Name authority. Say what each source governs and what happens when sources disagree.
- Close the loop. When an open question gets answered, move the answer into fixed decisions or rationale, and change the next action.
Read the thread before a new session, a handoff to a specialist, a follow-on project or anything that leaves the company. If the person or agent picking it up would still have to piece together why the work exists, a part is missing.
Where Supanova fits
Say the onboarding experiment runs as a Supanova project. The approved positioning guide goes in your workspace Codex, and the July interview synthesis is attached to the project itself. The atom drafting the outreach can draw text from both. The thread goes into the project brief your atoms work from, so the fixed segment, the role of each source and the growth lead's review point are there before any work starts.
Finished work comes back for your review, and that's one of the moments the thread gets updated. Say the growth lead sends the participant criteria back because two of the five companies look like enterprise accounts. Your team updates the thread so the fixed decision says what counts as mid-market. Keeping it current is part of directing the work.
Once the growth lead approves them, the outreach can go out. If a line in the finished draft points to the next job, you can highlight that passage and turn it into a new task or project.
Pick the next piece of work your team will hand off, to an agent or a colleague. Before it moves, write the six lines. If you can't fill in fixed decisions, that's the conversation to have first.
A transcript remembers the conversation. A Decision Thread remembers what now governs.
When you're ready to hand that work to an autonomous workforce, start at supanova.team.
Sources
- Lost in the Middle: How Language Models Use Long Contexts, Nelson F. Liu and colleagues, Transactions of the Association for Computational Linguistics, volume 12 (2024).
- LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory, Di Wu and colleagues, ICLR 2025.
- AI Risk Management Framework Core, National Institute of Standards and Technology, AI RMF 1.0 (2023).
- Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks, Microsoft Research (2024).
- A practical guide to building agents, OpenAI (2025).
- Effective harnesses for long-running agents, Anthropic Engineering (2025).