The attention layer — relevance computed over recorded work

Date: 2026-07-21 Status: blueprint (conclusion + staged design).

The underlying model: instrument every interaction as an automatic by-product of work, compute a per-path Degree-of-Interest (DOI) from weighted, decaying attention events, and surface ranked anchors — never asking anyone to curate them. Both the mechanism and the findings behind it bind this design.

The conclusion

The runtime already records everything: journals, task events, tool calls, reports, telemetry capture every interaction an agent has with the work. What has never been built is the relevance computation over those recordings. That is the attention layer, and it is the unifying design behind the Beam criterion: “semantic state-diff oversight” is ranked relevance computed over agent work.

Three doctrines bind this design as law:

1. The blind-spot doctrine (people call orientation aids useful and never maintain them by hand): every artifact of orientation must be an automatic by-product of work. Contenox already lives this — agents declared by filename, heartbeats riding tool calls, missions as records — and the attention layer must too: no curation, no tagging, no “mark as important”. Ranked relevance is computed, never asked for.
2. The metric lesson: navigation COUNT is not a tenable efficiency metric; the real signal is SCOPE — landing while touching fewer paths, and abandoning misleading paths faster. Step/token counts are the wrong efficiency metric for agents too. Scope is the signal: how few paths a unit touched to land, how fast it abandoned a wrong path. Corollary: derailment is a scope anomaly before it is anything else — the first real derailed unit (wandering $HOME instead of the repo) was detectable from its first two tool calls.
3. The falsifiability lesson: do not assume DOI-ranked oversight helps — measure it. Beam-over-LAN is an instrumented remote environment; recorded sessions, cohort comparison, and automated scoring are the study template the tool-eval harness already adopts. Rank-vs-flat review is A/B-able.

What it becomes, concretely (staged)

Stage 0 — the inward face: harness guidance counters (maintainer’s addition, 2026-07-21: “a tracker of some kind in the tool harness that would allow the LLM to be guided — like a simple counter”). The same attention signal, fed back to the AGENT live, through the tool-result envelope — the one channel every model reads. Deterministic per-session counters over (tool, path, args-fingerprint) yield: a repeat-call marker on identical calls (“[harness] 3rd identical list_dir of ’.’ this session”), a periodic scope line every N calls (“[harness] scope: 12 files in 4 dirs; mission cwd is …”), a revisit hint on heavy re-reads. The blind-spot premise applies to models verbatim — they cannot judge navigation value and will not curate their own navigation memory; the harness derives it. Rules: terse, fixed, clearly marked envelope (model-noise risk is real on weak models — the envelope is a tool-hardening surface and a future ModelProfile dial); NEVER a gate; counts reset per session; implemented as a ToolsRepo decorator so it wraps any tool provider without touching them.

Stage 1 — DOI over a mission’s work (rides the diff-review arc, ide-workflows.md Arc 1): per-path interest scores from the unit’s already- journaled interactions — reads, edits, tool dwell — weighted (edit > read > list), decayed, masked (one anchor per neighborhood, so a hot file does not drown out its whole directory). Surfaces: the changed-files list ordered by where the unit’s attention concentrated (review starts at the hot spot, not alphabetically); “concentrated in N files” badges on board rows.

Stage 2 — scope-anomaly detection (the derailment early-warning): a unit’s touched-path set diverges from its mission’s workspace/plan expectation → an attention-worthy condition in UnitStatus and the inbox, BEFORE failure or silence. Cheap: set arithmetic over the same aggregation. This converts the scope finding into the fleet’s most valuable alarm.

Stage 3 — DOI over the fleet for the operator: the overnight-skim navigation-cost problem is the single-session problem squared (20 units producing activity nobody can pre-judge). Rank inbox groups by accumulated relevance (blocker recency, plan-revision magnitude, scope anomaly, handover presence); anchors into long transcripts (jump to where attention/failure concentrated).

Stage 4 — landed-vs-planned, measured (closes the loop with the plan engine): the plan gives intent, the attention layer gives where work actually went; their diff IS the semantic state-diff review. The natural follow-up — mine derailed missions’ traces against landed ones — becomes an automatic fleet feature.

Design notes

• All inputs exist today (kernel journals, task events, tool-call events, reports); the layer is a consumer, never a new recording duty.
• Score mechanics: weighted event types, additive accumulation, decay per round, masking, relevance tiers — parameters are tunable and MUST be treated as hypotheses (lesson 3), not constants.
• Nothing here gates anything. The attention layer ranks and flags; envelopes gate. Rank is advice; the operator’s eyes stay the judge (exoskeleton, not autopilot).
• First implementation slice: the Stage-1 aggregation inside the diff-review arc’s Go endpoint (order-by-interest is one SQL/fold away from order-by-path), with edit-count ordering as the stub and the full DOI weights behind it.