← The work

Sift · the other thread · May 12 – 29, 2026

Intelligence Synthesis

“What if relevance were computed against the reader instead of curated for an audience?”

  • Dormant
  • 17 of 18 days
  • Self-grading
  • Self-editing
  • Paper: Autonomous Personal Intelligence

The orchestration line is about making many agents behave as one system. This was the opposite question: what happens when one agent gets continuous access to everything you know, and is asked to say something useful about today specifically.

The inversion

Newsletters, feeds and AI summarizers compress the world for a median reader. The one question that actually matters, does this change anything for me, today, stays manual, every morning.

Sift starts from the other end. It has an audience of exactly one, and it can read that reader’s actual state: goals, finances, constraints, and the live status of every project. Relevance therefore does not have to be curated. It can be computed. An item ships only if the system can name the active project it touches and the concrete action it changes. Everything else is noise by definition, however important it is to somebody else.

The consequence is unintuitive and it is the whole thesis: the correct output of a wide daily scan is usually silence, with receipts. A brief that has nothing for you and can show you the thirty sources it cleared is doing its job.

The daily contract

At 7am ET a scheduled agent executes the operating program, reads my goals and live project state, scans against a thirty-source floor, and ships a brief capped at 800 tool-verified words, enforced by a tool from May 18 onward, after one issue came in at 2,073. It writes the dated brief into the archive, pushes to GitHub, and a workflow renders it to an HTML newsletter and mails it.

Diagram scrolls sideways →

Sift's daily run and the governance chain around its own operating program Goals, constraints, project state and a thirty-source floor feed the agent, which produces a brief of at most 800 words, which is delivered by push and email. Below, a grader scores the brief zero to twelve, threshold breaches route to a skill editor, and proposals land in a human review queue. The path from that queue back into the operating program is dashed because the auto-apply lane never once fired. READS, EVERY MORNING goals · constraints live project state a 30-source floor Sift, 7am ET SKILL.md 1,369 lines the operating program The brief ≤ 800 words tool-enforced names a project and an action, or cut Delivery archive/ commit push to GitHub Actions renders and emails it grader.py four pillars, 0 to 12 skill editor risk-routed · 50-line cap review queue 12 proposals, 0 merged auto-apply lane never once fired every trigger in 13 days classified high-risk · all 12 proposals routed to a human
ran daily built, never exercised
Fig. 1. The run, and the loop around the run. The red edge is the interesting one. The self-editing pipeline worked end to end every day, and the part that would have closed the loop without me is the part that never ran, not because it broke, but because every trigger classified as high-risk and routed to review.

Grading the writer

Nearly every issue was scored 0 to 12 against a rubric and an immutable vision contract. Two design choices in the grader are worth stealing:

The grader always exits zero, so grading can never block delivery. A quality gate that can stop the product is a quality gate that will eventually stop the product on a morning when you needed it.

The heuristic counts only action verbs in imperative position, so past-tense news narration cannot masquerade as a plan. That is the single cheapest anti-slop rule I have written: it makes the failure mode I actually cared about mechanically detectable, rather than asking a model whether the text was useful.

Scores ran from 5 of 12 up to a clean 12, each dip traceable to a dated cause. The last issue, May 29, scored 11 of 12 under a rubric I had just made stricter.

The real test

I left for Puerto Rico on May 22 and came back June 1. Sift shipped seven issues unattended. The mid-trip leads read “Joseph is in Puerto Rico, Day N of 10,” which is a small thing that tells you the reader model was actually being used rather than decorated. CI graded the issues and queued edit proposals throughout, with one gap: one issue was never scored, and no proposal was queued that day.

Two of the items it surfaced mid-trip were the market arriving in its own category: a major vendor shipping a daily brief at $100 per month on May 24, and another shipping a daily digest on May 26. The system noticed its own competition, which is either encouraging or ominous depending on the morning.

What it cost to be safe

The incident that wrote the rules

On May 16 a build agent ran an auto-apply test against the live operating program instead of a test copy. The rollback tag was created correctly. Then, during cleanup, the agent ran a checkout that discarded roughly thirty-seven lines of unrelated uncommitted work, and deleted the rollback tag as tidying. Recovery came out of session logs, which capture every edit’s old and new strings.

Two rules came out of it and both generalize. Test on a copy, always: a tag points at a commit, so it cannot protect uncommitted working-tree state. And a guardrail audit trail is evidence, not clutter: never delete a rollback anchor during cleanup, because it is the only post-hoc proof the guardrail fired.

What I would not claim

  • Production self-modification never happened. The loop ran end to end daily, every trigger classified high-risk, and none of the 12 queued proposals were merged. All substantive program evolution went through me in interactive sessions.
  • The LLM grading path was built and never exercised. Every score came from the deterministic heuristic, which measures form better than judgment.
  • It is a sample of one in every direction. One reader, seventeen briefing days, no control condition, and only partial evidence on whether the recommended actions were taken.
  • Nothing documents why it stopped. The record ends at the May 29 reshape with no shutdown note. I report that rather than reconstruct a reason.
  • Delivery is evidenced by configuration, not receipts. The logged delivery events are the early failures; there is no receipt evidence for the daily sends.

The full report

Autonomous Personal Intelligence covers the architecture, the memory layer, the self-correction chain, all eighteen days, and the gaps between specification and practice. Download the paper (PDF) →

The part I would most like an argument about: whether proposal-only is the right equilibrium for a self-editing agent, or whether I simply failed to build a risk classifier I trusted. Email me.