All posts
Everything published here, newest first.
316 posts · showing 41–50
2026
10 posts-
Can You See the Raw Messages? The Question That Sorts Agent Frameworks
Every agent framework demos well. The one that survives your first 2am incident is the one that will hand you the exact bytes it sent to the model — and let you change them before they go.
-
A Two-Line Prompt Diff Is a Behavior Change. Here's the Regression Check It Needs.
A prompt edit has the blast radius of a code change and none of the safety rails — no compiler, no type error, no failing test. Six checks that turn a forty-second approval into a real regression gate.
-
Your Agent Called the Same Tool Three Times in a Row. Here's What the Loop Is Telling You.
A repeating agent isn't being stubborn — it's being consistent. Identical tool calls mean the context barely changed between turns, and there are only four reasons that happens.
-
A 5% Canary Mostly Samples Your Easy Tasks. Roll Agent Upgrades Out by Task Class.
Random traffic slices sample volume, but an agent's risk lives in the thin tail of long, tool-heavy, irreversible tasks. Here's how to build a rollout ladder that samples the danger instead.
-
Your Agent's Dry Run Doesn't Bind the Real Run — Confirm the Plan Hash, Not the Intent
A human approves a preview; the agent then executes a freshly computed call. Everything that shifts in between is unguarded — and the fix is a redeemable plan object, not a sterner confirmation prompt.
-
When the Context Window Fills, Cut in This Order — and Leave a Stub Behind
Most agents evict context by dropping the oldest tokens, which reliably deletes the nine words that mattered while keeping a 4,000-token HTML dump nobody needs. Here is the order to cut in, and why every eviction should leave a stub behind.
-
"Done" Is Not a Handback: What a Worker Agent Should Actually Return to Its Supervisor
A worker agent reporting "completed" tells its supervisor nothing it can verify. Here is the handback shape that makes a fabricated result catchable by code instead of by a second model's opinion.
-
Your Tool Returned 4,000 Rows and the Agent Needed Three: Truncate, Paginate, or Summarize?
Most tool endpoints were written for a caller that can loop and discard. An agent reads the whole response in one gulp and keeps it forever — and the three ways of shrinking that response fail in completely different ways.
-
One 300-Second Timeout Tells You Nothing. Agents Need a Deadline Budget.
A run that dies at its ceiling with no attribution tells you nothing about which part was slow — and pushes you toward the laziest fix, raising the ceiling. Three nested clocks, an inherited remainder, and a reserved tail slice turn timeouts into a signal you can actually read.
-
Untrusted Text Doesn't Stay Untrusted: Where Provenance Gets Lost in an Agent Pipeline
An injection scanner that fires on the raw email won't fire on the tidy summary of it. Untrusted content gets laundered clean every time an agent rewrites it — here's where the taint disappears, and how to make it survive.