diff options
| -rw-r--r-- | gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi | 7 | ||||
| -rw-r--r-- | gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi.tpl | 6 | ||||
| -rw-r--r-- | gemfeed/atom.xml | 9 |
3 files changed, 18 insertions, 4 deletions
diff --git a/gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi b/gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi index 2fc6bbc1..d4c77e4b 100644 --- a/gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi +++ b/gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi @@ -26,6 +26,7 @@ I let Ampcode autonomously complete 48 Taskwarrior tasks on my eBPF project in a * ⇢ ⇢ ⇢ 4-annotate-update-task.md — progress tracking * ⇢ ⇢ ⇢ 5-review-overview-tasks.md — picking the next task * ⇢ ⇢ The reflection and review loop +* ⇢ ⇢ Code review: human spot-check at the end * ⇢ ⇢ Measurable results * ⇢ ⇢ A real bug found by the review loop * ⇢ ⇢ Gotchas and lessons learned @@ -414,6 +415,10 @@ Having instructed in the skill for the agent to reflect on its own implementatio The sub-agent reviews consistently caught things the main agent missed — tests that only asserted on mocks, missing edge cases, and even a real bug. Without the dual review loop, the agent tends to write tests that look correct but do not actually exercise real behavior. +## Code review: human spot-check at the end + +On top of the agent's self-reflection and the two sub-agent reviews per task, I reviewed the produced outcome at the end. I did not read through all 5k lines one by one. Instead I looked for repeating patterns across the test files and cherry-picked a few scenarios — for example one integration test from the open/close family, one from the rename/link family, and one negative test — and went through those in detail manually. That was enough to satisfy me that the workflow had produced consistent, runnable tests and that the whole pipeline (task → implement → self-review → sub-agent review → fix → second review → commit) was working as intended. + ## Measurable results Here is what one day of autonomous Ampcode work produced: @@ -422,7 +427,7 @@ Here is what one day of autonomous Ampcode work produced: * 48 Taskwarrior tasks completed * 47 git commits * 87 files changed -* 12,012 lines added, 1,543 removed +* ~5,000 lines added, ~500 removed * 18 integration test files * 15 workload scenario files (one per syscall category) * 93 test scenarios total (happy-path and negative) diff --git a/gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi.tpl b/gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi.tpl index a067d40d..c62e92e1 100644 --- a/gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi.tpl +++ b/gemfeed/2026-02-22-taskwarrior-autonomous-agent-loop.gmi.tpl @@ -390,6 +390,10 @@ Having instructed in the skill for the agent to reflect on its own implementatio The sub-agent reviews consistently caught things the main agent missed — tests that only asserted on mocks, missing edge cases, and even a real bug. Without the dual review loop, the agent tends to write tests that look correct but do not actually exercise real behavior. +## Code review: human spot-check at the end + +On top of the agent's self-reflection and the two sub-agent reviews per task, I reviewed the produced outcome at the end. I did not read through all 5k lines one by one. Instead I looked for repeating patterns across the test files and cherry-picked a few scenarios — for example one integration test from the open/close family, one from the rename/link family, and one negative test — and went through those in detail manually. That was enough to satisfy me that the workflow had produced consistent, runnable tests and that the whole pipeline (task → implement → self-review → sub-agent review → fix → second review → commit) was working as intended. + ## Measurable results Here is what one day of autonomous Ampcode work produced: @@ -398,7 +402,7 @@ Here is what one day of autonomous Ampcode work produced: * 48 Taskwarrior tasks completed * 47 git commits * 87 files changed -* 12,012 lines added, 1,543 removed +* ~5,000 lines added, ~500 removed * 18 integration test files * 15 workload scenario files (one per syscall category) * 93 test scenarios total (happy-path and negative) diff --git a/gemfeed/atom.xml b/gemfeed/atom.xml index 4b89d2cf..1b028d67 100644 --- a/gemfeed/atom.xml +++ b/gemfeed/atom.xml @@ -1,6 +1,6 @@ <?xml version="1.0" encoding="utf-8"?> <feed xmlns="http://www.w3.org/2005/Atom"> - <updated>2026-02-21T23:24:01+02:00</updated> + <updated>2026-02-21T23:38:45+02:00</updated> <title>foo.zone feed</title> <subtitle>To be in the .zone!</subtitle> <link href="gemini://foo.zone/gemfeed/atom.xml" rel="self" /> @@ -47,6 +47,7 @@ <li>⇢ ⇢ <a href='#4-annotate-update-taskmd--progress-tracking'>4-annotate-update-task.md — progress tracking</a></li> <li>⇢ ⇢ <a href='#5-review-overview-tasksmd--picking-the-next-task'>5-review-overview-tasks.md — picking the next task</a></li> <li>⇢ <a href='#the-reflection-and-review-loop'>The reflection and review loop</a></li> +<li>⇢ <a href='#code-review-human-spot-check-at-the-end'>Code review: human spot-check at the end</a></li> <li>⇢ <a href='#measurable-results'>Measurable results</a></li> <li>⇢ <a href='#a-real-bug-found-by-the-review-loop'>A real bug found by the review loop</a></li> <li>⇢ <a href='#gotchas-and-lessons-learned'>Gotchas and lessons learned</a></li> @@ -437,6 +438,10 @@ tasks are in progress, show the next actionable (READY) task: <br /> <span>The sub-agent reviews consistently caught things the main agent missed — tests that only asserted on mocks, missing edge cases, and even a real bug. Without the dual review loop, the agent tends to write tests that look correct but do not actually exercise real behavior.</span><br /> <br /> +<h2 style='display: inline' id='code-review-human-spot-check-at-the-end'>Code review: human spot-check at the end</h2><br /> +<br /> +<span>On top of the agent's self-reflection and the two sub-agent reviews per task, I reviewed the produced outcome at the end. I did not read through all 5k lines one by one. Instead I looked for repeating patterns across the test files and cherry-picked a few scenarios — for example one integration test from the open/close family, one from the rename/link family, and one negative test — and went through those in detail manually. That was enough to satisfy me that the workflow had produced consistent, runnable tests and that the whole pipeline (task → implement → self-review → sub-agent review → fix → second review → commit) was working as intended.</span><br /> +<br /> <h2 style='display: inline' id='measurable-results'>Measurable results</h2><br /> <br /> <span>Here is what one day of autonomous Ampcode work produced:</span><br /> @@ -446,7 +451,7 @@ tasks are in progress, show the next actionable (READY) task: <li>48 Taskwarrior tasks completed</li> <li>47 git commits</li> <li>87 files changed</li> -<li>12,012 lines added, 1,543 removed</li> +<li>~5,000 lines added, ~500 removed</li> <li>18 integration test files</li> <li>15 workload scenario files (one per syscall category)</li> <li>93 test scenarios total (happy-path and negative)</li> |