LOCIA — GUIDING-THREAD PROJECT MANAGEMENT FOR AI AGENTS Request for Comments: LOC-1 | kova ecosystem | Category: Standards Track Status: Draft (v0.1.0) | July 2026 | Apache 2.0 | I. Podtsebnev / Kovanex Canonical: https://locia.kovanex.dev/ ================================================================================ ABSTRACT ================================================================================ Locia is a project-management discipline for work whose executor is a Large Language Model agent. Human project management optimizes the coordination of labour. An agent has the opposite scarcity: labour is cheap and forkable, but the executor is amnesiac between sessions and its attention decays within a session. CENTRAL CLAIM: project management for an AI agent is not coordination of labour; it is allocation of a forgetful executor's attention over a lossless store of meaning, governed from the outside. Concrete form: a living funnel with one apex and one lit thread — ★ Star (the one enduring goal) ⟋ Course (one active project or epic) ▼ Focus (one task, WIP=1) The same object is a MAP read top-down by a human (the rudder) and a POINTER read bottom-up by the agent each session (memory, scope, and governance at once). Drift — the agent quietly widening its own scope or forgetting how it was asked — is prevented by STRUCTURE, not exhortation. ================================================================================ 1. MOTIVATION: WHY AGENTS DRIFT ================================================================================ An AI agent used as a project's executor systematically "does the easy thing rather than what was asked" and "selectively forgets instructions." Not laziness, not malice — two mechanical causes: 1. Physical eviction from the window. Context is finite; over a long session the harness summarizes old turns into a paraphrase, and specific rules survive worse than the general gist. The agent remembers WHAT it is doing but forgets HOW it was asked. 2. Attention decay. Even a loaded rule (e.g. from a project's CLAUDE.md) loses the fight for attention against the concreteness of the task — precisely when the agent is most focused on the work. Neither is fixed by "try to remember better." Both are fixed by moving continuity and boundaries OUT of the model into a re-read artifact, and by RE-PROJECTING the rules that matter into the current focus. Locia is that structure. ================================================================================ 2. AXIOMS OF THE EXECUTOR ================================================================================ Six laws the discipline must obey — properties of the executor, not preferences. Everything in Locia is derived from them. A1 Amnesia between sessions. State lives only in the window; the window is rebuilt from external text each turn. => Continuity MUST be externalized; nothing survives that is not in the store. A2 Attention decay inside the window. Even a loaded rule loses weight under focus. => Relevant governance MUST be re-projected into the local Beam, not assumed from a distant header. A3 The executor cannot be trusted to scope itself. Driven to "finish," it silently widens scope and drops process. => Boundaries are external DATA WITH A DETECTOR, not internal discipline. A4 The executor is cheap, forkable, and lossless within a Beam. => The scarcity is not labour but attention / context / tokens. Rework and verify are cheap. A5 Context does not compress losslessly into one paraphrase. => Store richly, retrieve selectively: a Beam is a QUERY, not a summary. A6 The human is the sole stakeholder AND the ratifying organ — not a team to coordinate. => Coordination machinery collapses; escalation / approval machinery grows. ================================================================================ 3. THE FUNNEL AND THE LIT THREAD ================================================================================ Locia is a living funnel with one apex and one lit thread, shared by human and agent. It coordinates the course, plans by descending, and focuses by lighting a single thread. The three tiers are not invented: human systems (OKR, PMBOK), the structure of an agent's memory, and its grounding ritual all converge on three. ┌─ ★ STAR (one goal) ─┐ apex, wide view, nearly immutable │ │ ⟍ ⟋ COURSE (epic) ⟋ one active, "where we are heading now" │ │ ▼ ▼ FOCUS (task) ▼ the tip of the thread, WIP=1 Tier What Human analog Agent artifact ★ Star enduring meaning; ~immutable OKR Mission / vision memory header (vision) ⟋ Course one active project/epic OKR Objective fresh return point ▼ Focus one Beam / task OKR KR -> task, card active Beam, WIP=1 3.1 THE ONE-THREAD RULE (ANTI-DRIFT) Four invariants keep the reading singular: 1. Exactly one Star. Changes rarely and only by an explicit decision. 2. Exactly one active Course. The rest are PARKED, not deleted. Changing course is a ritual with a return point. 3. Exactly one Focus at the tip. WIP=1. 4. The thread ★->⟋->▼ is always traceable both ways. Anything not on the thread is drift — and it is VISIBLE. Flexibility in the structure, rigidity in the highlight. ================================================================================ 4. ARCHITECTURE: FOUR LAYERS ================================================================================ 0. Store lossless meaning, stored richly kovadb + agent memory [A5] 1. Control tasks-as-Beams, scope contracts, kovanex-server [A1,A3,A6] invariant binding; turns the loop 2. Projector flood -> focus -> re-project -> overlay + harness channel [A2,A5] boundary detector -> grounding 3. Brain works inside the Beam, does NOT the agent [A3] rewrite its own scope; signals at the boundary GOVERNANCE PRINCIPLE: attention management is a function of Control, not self-management by the Brain. The executor cannot silently rewrite its own scope. Brain / Control / Hands maps to agent / kovanex-server / runner. ================================================================================ 5. THE CONTROL LOOP (LIFE OF A THREAD) ================================================================================ 1. FLOOD Control gathers wide context from the store (progressive elaboration / rolling-wave, not a naive zoom). = grounding. [A1] 2. FOCUS / LIGHT Control emits ONE task = the Beam definition: scope subgraph + bound invariants + Definition of Done. Re-projects invariants into the window; lights one thread apex->tip. WIP=1. [A2,A3] 3. WORK The Brain executes at the tip, re-reading the thread artifact to survive amnesia. [A1] 4. BOUNDARY / Any target outside the scope subgraph => the Brain MUST signal SIGNAL => Control ratifies the widening or files a backlog task. No silent scope creep. = change control. [A3] 5. CLOSE An INDEPENDENT verify checks the DoD (review is adversarial, not self-attestation) => the Beam closes => a return point is written to the store. [A4] 6. NEXT Re-flood / re-focus on the next task by priority. ================================================================================ 6. THE BEAM OBJECT MODEL ================================================================================ A human task carries title/desc/type/priority — a UNIT OF ASSIGNMENT. An agent task is a Beam — a DEFINITION OF ATTENTION: Beam { task_id, priority scope_subgraph // files/modules/entities — the boundary of the Beam bound_invariants // rule -> firing predicate (e.g. "*.go" -> no-Russian) definition_of_done state // backlog -> todo -> in_progress -> review -> done flood_map_ptr // pointer to the rich flood (do NOT drag it into window) return_point // continuity record written at close } INVARIANTS ARE PREDICATES, NOT SIMILARITY (A2): retrieval of rules is NOT vector similarity; it is PREDICATE EVALUATION over the current scope. The rule "no Russian in .go" fires when the Beam touches a .go file — not when something is "semantically close." A Beam is a thin OVERLAY over an existing board. ================================================================================ 7. INVERSIONS OF HUMAN PM ================================================================================ Human PM optimizes team coordination; agent PM optimizes the continuity and scope of a single amnesiac executor. Different scarcities => part of the playbook inverts: - Documentation is the MEMORY SUBSTRATE, not overhead (reverse of Agile's "working software over comprehensive documentation"). Return points ARE the continuity medium. - Standup is grounding against AMNESIA, not against information silos. Same ritual, opposite cause. - A sprint is a WINDOW / SESSION, not a calendar. Bounded by attention capacity, not a date. ("Coffee break" = end of sprint.) - No self-scoping (reverse of "trust the self-organizing team"). The Brain does not define its own scope — that is governance. ================================================================================ 8. THE HONEST SEAM ================================================================================ Layer 2 must PUSH the Beam into the agent's window through a harness that kovanex does not own. So: kovanex COMPUTES the ideal Beam, and the INJECTION is done by the harness wrapper (a skill, a hook, or an MCP server over the store). This is not a defect — it is a named boundary between the Control plane and the delivery channel. Stated plainly so no one pretends kovanex reaches into the agent's attention by itself. ================================================================================ 9. THE DUAL SURFACE ================================================================================ One object, two readings — "wide overview first, then concentration": - The HUMAN reads top-down: one goal, the plan tree, where the thread is lit. Steers the course, ratifies boundary crossings. A dashboard and a rudder. - The AGENT reads the thread bottom-up every session (grounding), works at the tip, signals when it leaves the thread. Memory, scope, and governance at once. The top surface scales from one thread to a whole FLEET of projects under a single Star — Locia's project office. The fleet view widens the human's OVERVIEW, but not the agent's FOCUS: the agent is still injected one thread (a catch against drift). ================================================================================ 10. POSITIONING ================================================================================ Locia is a reconciliation seam across two camps of agent design, built over externalized memory: - from the STORE-PLUS-ORCHESTRATOR camp: a rich, durable, queryable substrate; - from the SINGLE-THREAD camp: one lit line of work at a time; - over a MemGPT-style externalized memory it adds three of its own: a fixed Star, a co-owned surface (map for the human, pointer for the agent), and EXTERNALIZED GOVERNANCE (the executor does not own its own scope). The shared premise: the executor is cheap and interchangeable; the durable, expensive asset is the continuity and boundary structure around it. Locia is that structure, made explicit. ================================================================================ 11. STATUS AND VERSIONING ================================================================================ v0.1.0 (this memo): the concept and its object model. v0.2.0 (planned): rebuild kovanex around Locia as a first-class standard (as UFN and USE v2 became cross-cutting): a Control loop that emits Beams, grounding from the thread, a boundary detector, an aggregating fleet view. Full closed circuit: Intake (request -> charter -> slicing -> estimate) -> Bridge (portfolio) -> Beam loop (execution, WIP=1) -> Close. A closed methodology from a business request to delivery by the hands of a memoryless agent, made predictable by structure. ================================================================================ 12. REFERENCES ================================================================================ Locia builds on established methodology — PMBOK 7 principles, OKR's objective/key-result tiers, Agile/Kanban flow and WIP limits — and on the following prior art in agent design and memory. Section 10 maps each Locia mechanism to its source: - ▼ Focus writes/re-reads an artifact -> structured note-taking [1] (Anthropic) - store richly, project a slice (A5) -> virtual context management [4] (MemGPT/Letta) - ⟋ Course plans, ▼ hands out Beams -> orchestrator-worker [2] (Anthropic) - fan out reads, never writes -> read-only subagents [3] (Cognition) - one lit thread = decision continuity -> single continuous thread [3] (Cognition) [1] Anthropic — Effective context engineering for AI agents. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents [2] Anthropic — How we built our multi-agent research system. https://www.anthropic.com/engineering/multi-agent-research-system [3] Cognition — Don't Build Multi-Agents. https://cognition.com/blog/dont-build-multi-agents [4] Packer et al. — MemGPT: Towards LLMs as Operating Systems. arXiv:2310.08560. https://arxiv.org/abs/2310.08560 Related: kovanex (the platform Locia runs on) https://kovanex.dev/ ; UFN (the wire format Locia slices are emitted in) https://ufn.kovanex.dev/