Back to articles

Tag archive

#contextengineering

T
Oct 23, 2026

The agent drifts from your requirement while the context fills up

You hand an agent a short requirement, it makes ten or twelve tool calls, and the final diff quietly solves the adjacent problem instead of the one you asked for. The window fills with fetched files, command output, and trial edits; recent tokens weigh more than the original brief; and nothing forces the agent to re-check acceptance mid-task. This article names the drift patterns (optimize-the-last-output, summary-takes-over, scope-creep-by-convenience, stale-tool-log, lost-initial-criteria), then gives the cheap anchors: a pinned acceptance block, a periodic restate-and-compare check, pruning stale tool logs, and a written handoff between sessions.

Oct 23, 20263 min read1 reactions0 comments
T
Oct 22, 2026

The agent wrote its own test input, watched it pass, and called the edge case closed

An agent fixes an edge case from a messy real input, then verifies the fix against a tidy synthetic record it wrote itself. The synthetic input parses cleanly, the check goes green, and the bug is declared closed, while the original malformed payload that triggered the incident never gets replayed. This article covers why self-authored inputs grade themselves, the telltale signs in the thread, and the cheap gate of keeping the real failing payload as the permanent regression case instead of inventing a friendlier substitute.

Oct 22, 20264 min read0 reactions0 comments
T
Oct 25, 2026

The agent kept implementing the approach we already rejected because it never left the context window

When a developer redirects an agent away from an explored approach, the rejected branch often stays in the context window and keeps getting implemented. This piece explains why an agent remembers proposals without remembering their status, what extra abstractions and unrequested tests ship as a result, and how to keep an explicit active-constraint list separate from dropped exploration so a rejected idea stops steering the code.

Oct 25, 20263 min read0 reactions1 comments
A
Oct 25, 2026

After three hours the agent defended the wrong model it invented, because the session remembered the assumption better than the facts

In a long agent session, the first plausible guess becomes the working model, and later corrections attached at the bottom read as exceptions instead of a rewrite. The session then defends the invented convention, edits against the wrong frame, and misses contradicting evidence. This article explains why long chat history anchors on early assumptions, what breaks when a correction arrives late, and how a compact verified-facts brief, periodic session reset, and re-derivation from source files keep the agent from hardening around a guess.

Oct 25, 20263 min read0 reactions1 comments