Working with AI on analysis · 6/7
If context still fits, should we keep adding to it?
Suppose an agent has a 1M context window and compacts automatically at 800k. These are hypothetical numbers, not the specifications or thresholds of a particular product. My question is whether the current analysis really needs to keep carrying all 800k tokens.
I want the progress of the work to help determine what stays active and what can be saved for reference.
Capacity and analytical progress are different signals
After working with an agent for a while, the conversation usually contains several SQL versions, results based on different denominators, tentative explanations, and successive corrections.
For event participation, we might start with all active users, switch to users who meet the eligibility criteria, and finally adopt users who have unlocked the event entrance.
Those discussions matter. Once the denominator is settled and we move to segmentation, do we still need every trial and conversation turn in every input?
Beyond token consumption, I worry about mixing the current definition with discarded alternatives. We just agreed not to use that denominator, and the agent brings it back. Apparently we were only exchanging polite opinions.
Summarizing conversation near a capacity limit is a common approach. But remaining context capacity and whether a phase of analysis can be narrowed are different questions. They can even conflict: an early decision goes uncompressed, then a large investigation gets interrupted by compaction and has to retrieve its data again.
Our decisions already provide signals for compaction
Things we naturally say during analysis are useful signals:
| Collaboration signal | What to preserve | A distinction that must remain visible |
|---|---|---|
| “This denominator is confirmed; move on” | The adopted definition, rationale, and scope | Adopting a definition does not verify all results |
| “Stop this path for now; the sample is too small” | The method, stopping reason, evidence location, and conditions for reopening | Stopped does not mean successfully completed |
| “Keep this result; we will compare it with another group” | The result and its location, plus the outstanding comparison | A partial result is not a final conclusion |
We are already making these judgments. Turning them into usable state at that moment reduces the need to repeatedly infer what was merely proposed, what was accepted, and what still needs validation.
I have been calling this “natural compaction”: letting context narrow in stages as the analysis progresses, based on people’s choices. This is my name for a way of working, rather than a claim about a new compression algorithm.
A general summary must infer importance from existing records. An analysis workflow has another signal available: we are actively reviewing, validating, and adjusting methods. Those choices can guide what gets retained.
Sometimes the right choice is not to narrow yet. If a person says to keep something open, it should remain unresolved. A polished summary must not quietly settle the question.
“I agree to proceed this way” and “this result has been validated” must stay distinct. Adopted definitions, checks that passed, and open doubts need their evidence references. Otherwise, after a few summaries, everything may look confirmed.
Carry less now, but keep a path back to the details
Compaction does not mean we will never need the omitted details again.
A manager may ask: “How did you decide the sample was too small?” A note saying “insufficient sample; stopped” is not enough. We still need the SQL, sample distribution, and decision criterion.
A summary usually cannot reconstruct all the details it omitted. The workflow needs both an active context and an external record:
- Current context: the adopted Plan and definitions, necessary reasons, unresolved questions, and evidence locations.
- Workspace: Plan versions, queries, results, and review records, connected through lineage so the available details can be retrieved.
We do not need to carry an entire old Plan for a stopped path into every turn. We do need to know why it stopped and where the evidence lives, so new data can prompt a sensible decision about reopening it.
This is the direction I am working toward in Lantide Data: use review, execution, validation, and iteration as natural points for organizing context. Save state based on confirmed choices, then retrieve what the next task needs.
Summaries remain lossy, and organizing and retrieving information have costs. The opportunity is to let the work itself help decide what needs to stay active, while keeping a route back when the question changes.
Further reading
Try this way of working in Lantide Data: Download Lantide Data.
Previous：How Can a Team’s Analysis Become Experience That People and Agents Can Reuse? · Next：Beyond Document Versions: Preserving Analytical Turning Points with Change Context