Skip to content

Repository files navigation

Shadow-Workstream Discovery for Autonomous LLM Agent Fleets

Defensive publication + clean-room reference implementation. Establishing dated, enabling public prior art over a mechanism that mines raw agent-fleet transcripts to discover untracked work initiatives — work that emerged rather than was assigned — using negative-set-conditioned LLM extraction and seen-count reinforcement.

Publication date: 2026-07-03 · Author: Gustavo Assuncao, PhD · Publisher: Gus IT LLC (Florida, USA)


What this is

When you run a fleet of autonomous LLM engineering agents, work is constantly initiated without ever being chartered. An agent chases a root cause and quietly starts a substantive multi-hour initiative; the only durable trace lives in its raw session transcript, never in your tracker. Your portfolio, burndown, and capacity models go systematically blind to a growing slice of what the fleet actually does.

This repository discloses — as public prior art — a mechanism that fixes that. On a fixed cadence it:

  1. Builds a negative set from the union of your tracked project-of-record and your existing proposals ledger, normalized to canonical slugs.
  2. Condenses a sliding window of raw per-agent session transcripts into a bounded, recency-biased, per-agent-quota'd corpus (with a channel to inject orchestrator work that isn't in the transcript store).
  3. Prompts an LLM with the corpus and the negative set, instructing it to conservatively extract only new, substantive initiatives as strict JSON — each with an owner and an evidence pointer.
  4. Reconciles survivors into a slug-keyed proposals ledger: first sighting inserts, re-sighting bumps a seen_count and last_seen_at. Recurring signal accumulates confidence; one-off chatter decays out of the sliding window on its own.

The result is a continuously-refreshed, ranked list of candidate untracked initiatives — named, owned, evidenced — offered to your portfolio process for promotion, dismissal, or merge. It is a portfolio dark-matter detector for autonomous fleets.

Why it's published (prior-art purpose)

The individual ingredients (cron, LLM extraction, negative prompting, slug dedup, upserts, sliding windows) are commodity. The combination — reconstructing named/owned/evidenced initiatives from raw agent transcripts, subtracting the project-of-record, and using cross-run recurrence as a self-maintaining confidence signal to catch emergent work — is a narrow, useful, non-obvious aggregate: exactly the kind of thing that gets patented. We publish it openly so it stays free for everyone to practice. This is a defensive publication: it is intended to constitute enabling public prior art as of 2026-07-03.

Prior-art risk is assessed MEDIUM; see docs/PRIOR-ART.md.

What's here

Path Contents
DEFENSIVE-PUBLICATION.md The full ~5,000-word disclosure: mechanism, data model, worked example, prior-art delta, and defensive claims. Start here.
docs/FIGURES.md Mermaid diagrams (pipeline, condensation, reinforcement, reconciliation, lifecycle).
docs/PRIOR-ART.md Real references, delta table, and the honest novelty nub.
docs/OPEN-SOURCE-APP.md How the reference maps to a deployable open-source app + AKS/CronJob sketch.
src/ Clean-room, dependency-free, offline-runnable Node.js (ESM) reference.

Run the reference

No dependencies, no network, no credentials. Requires only Node.js (ESM).

cd src
node example.js

example.js reproduces a deterministic three-run trace: a durable recurring initiative climbing to seen_count = 3, an emergent-then-recurring initiative climbing to 2, and a one-off frozen at 1 and going stale — demonstrating both negative-set exclusion and seen-count reinforcement. The LLM is an injectable stub; the datastore is an in-memory store implementing insert/bump upsert semantics.

Reference implementation notes

The reference is intentionally decomposed along the seams a real operator must fill: an injectable model function (llm(prompt) -> text) and an injectable datastore (tracker, intake, transcripts). It contains no proprietary code, no real provider calls, and no secrets. Empirically-tuned production constants (corpus caps, per-agent quotas, item limits, thresholds) are not disclosed — they appear as named config with illustrative defaults and are marked [WITHHELD — trade secret]. The mechanism is fully enabled without them.

License

Dual-licensed:

  • AGPL-3.0-or-later — free/open use under the terms of the GNU Affero General Public License, version 3 or later.
  • Commercial license — for organizations that cannot accept AGPL terms, a commercial license is available from Gus IT LLC. Contact gus@gusit.de.

The defensive-publication text is released to the public to serve as prior art. Third-party names/marks referenced are the property of their owners. Nothing here is legal advice.


© 2026 Gus IT LLC (Florida, USA). Author: Gustavo Assuncao, PhD.

About

A periodic pipeline that reconciles raw autonomous-agent session transcripts against the project-of-record set to surface work that emerged rather than was assigned, with recurrence-reinforced confidence.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages