Watch & compare
Context engineering
Different ways to see the same idea. Play one cut at a time, then return to the guide to try it yourself.
Why we recommend Keep the contract
Keep the contract is the strongest chapter entry because it shows a lost requirement cause a concrete compatibility failure, then requires a fresh fixture check after the handoff is repaired. Baseline complements it by separating log compaction from caller retrieval while retaining a constraint. Fable adds the most useful iterative retrieval loop, but its denser rows and smaller details take more effort to follow at phone width.
Revised original · 25 sec
Compact logs, retain the constraint
Summarize redundant history while retaining requirements, then retrieve the evidence needed for the next model invocation. A generated patch still needs checking.
Read the video sequence
- 0.0–7.7s: KEEP API, the null-crash task and api.ts:42 trace occupy an illustrative 8k-token context. Three repeated logs raise the displayed count to 10k, 13k and 15k of 16k.
- 7.7–11.0s: Redundant logs fade and become summary: null input. The unchanged constraint, task and trace remain; with the summary the illustrative total is 9k. The constraint and trace are highlighted.
- 12.2–15.3s: A callers? request reaches Repository. caller.ts returns and is recorded as caller.ts:18; the context rises to 11k.
- 17.3–22.9s: The selected context reaches Model. After processing, an amber Patch? candidate appears. No test or correctness verdict is shown.
- 23.0–25.0s: The complete scene fades to black; the exact decoded first and last frames match.
About this cut
Production: Original prototype; model not recorded.
Revision: gpt-6-astra comparative review.
Token counts and 16k capacity are illustrative, not a named model configuration or measured compression ratio.
The example shows a successful selection of retained details; summaries can omit requirements. Inspect the summary against its sources.
The final Patch? is a candidate, not a tested result.
Audio was analyzed, not heard.
Recommended · Astra · 30 sec
Keep the contract
Preserve compatibility constraints across agent handoffs, reopen their source, and rerun the matching check before accepting a patch.
Read the video sequence
- In this illustrative invoice task, the agent receives a goal to fix a timeout and a requirement to keep the v1 response field total_cents. A compatible patch passes the v1 fixture, establishing the normal case.
- A handoff keeps the timeout goal but drops the compatibility requirement. The missing row turns red and empties. This shows one possible loss of context, not an inevitable result of summarization.
- A second agent session uses the incomplete handoff and returns the replacement field amount. The same v1 fixture now fails, making the incompatible change visible.
- The v1 contract is reopened. Its total_cents requirement is loaded back into the checked handoff. The fixture remains failed until the patch is repaired and checked again.
- The agent reads the repaired handoff, restores total_cents in its patch, and reruns the fixture. It passes this check, while human review remains pending. The outcome is illustrative: context does not guarantee compliance or enforce permissions.
About this cut
Production: gpt-6-astra.
The invoice example is an illustrative scenario, not a recorded agent run or a measured failure rate.
A missing requirement may be rediscovered, and restoring context does not guarantee compliance.
The v1 fixture checks only its covered behavior; human review remains pending.
Written or persistent instructions do not enforce permissions; authorization is outside this lesson.
Audio was analyzed for peak, timing, and spectral events, but was not physically listened to.
This is an independently authored alternative; no comparative winner is claimed.
Alternative · Fable · 26 sec
Load the trace and one module, fetch the rest on demand
Retrieve the failing trace and the module it identifies, then follow dependencies revealed by inspection. Stale notes can steer a candidate edit; fewer tokens do not establish a speedup or a correct patch.
Read the video sequence
- 0.0–4.8s: In this retrieval example, the null-crash task is present before the trace and all five repository files are loaded. The illustrative meter rises from 1k to 14k of 16k. High occupancy turns amber; the window is not exceeded.
- 5.4–11.9s: Context reaches the model, which pulses while working. At 7.3s the old notes row exposes old: edit db.ts. An amber db.ts? candidate leaves the model at 9.3s and lands at 9.8s. No test has run; the candidate reflects the stale note.
- 12.0–12.8s: The context is reset to the task alone and the previous candidate clears. The meter returns to 1k.
- 13.2–15.7s: The trace is loaded first; its api.ts module is selected and retrieved. The illustrated context settles at about 4k.
- 16.1–19.9s: After reading the partial context, the model requests users.ts?; the repository row lights and the file returns. The context grows to about 7k.
- 20.1–25.8s: Another model invocation emits the amber api.ts:42? candidate at 21.7–22.2s. The proposal holds until 24.7s, then the stage fades to the title-only opening. No latency number, test result or correctness tick appears.
About this cut
Production: anthropic/claude-fable-5-1.
Revision: gpt-6-astra comparative review.
This is a schematic contrast, not a recorded agent run. Stale evidence may mislead, but a large context can contain essential information and selective retrieval can miss it.
Token counts and the 16k window are illustrative. Runtime retrieval can add tool calls and latency; caching can reduce repeated-input processing cost. There is no measured speedup here.
Both output badges are candidate locations, not patches checked by a fixture. The trace alone does not prove the fault lies in the named file.
One follow-up retrieval is shown; real agents may need several and should preserve applicable constraints throughout.
Material placed in context informs the model; it does not enforce permissions. Authorization belongs in the harness and tools.
Audio was analyzed for peak, timing and spectral events, not heard.
These animations use sound effects rather than narration. The written guide explains the mechanism and its limits. Model names describe how a cut was made, not a quality ranking.
What the review found
- Astra 6.3–13.8s makes the missing total_cents constraint visible before amount fails the same v1 fixture; baseline 7.7–11s shows only successful retention, and fable 7.3–9.8s shows stale advice influencing an unverified candidate.
- Baseline 7.7–15.3s most clearly separates compaction from retrieval: repeated logs become a summary, then caller.ts is fetched. Corrected illustrative totals of 9k and 11k preserve the original 8k packet consistently.
- Fable 16.1–19.9s uniquely shows an initial model invocation discover the need for users.ts and request more evidence. This is a useful complement to the single retrieval in baseline 12.2–15.3s and contract reinspection in astra 16–21s.
- Astra 21–26s reruns the fixture before PASS and leaves Human review pending. Baseline 19.1–22.9s and fable 21.7–24.7s now end on amber candidates because neither demonstrates a test; no cut establishes a timeout improvement.
- At 350px, astra 10.3–25.5s has the clearest labels and least reading competition. Baseline 8.5–15.5s is also readable; fable 3.5–9.8s remains denser despite enlarged rows and corrected packet routes, and benefits from replay.
- Fable 6.1–22.2s no longer presents invented 41s/9s timing as evidence. Baseline holds its candidate from 19.1s before fading at 23s; astra is longer at 30s but gives the loss, failure and recovery distinct reading time.