Intent, Custody, Trajectory, and Proof — Accountable Agentic Execution

Status Pre-results technical report. Controlled metrics remain TBD. Synthetic fixtures are schema tests; a separately labeled git-derived Decapod snapshot is observational, not treatment evidence.

Abstract

Large language model (LLM) agents are transitioning from conversational programming assistants to autonomous executors capable of editing codebases, running build tools, executing tests, and managing multi-file software deployments. Many coding-agent workflows still rely heavily on prompt transcripts and diffs as durable records. As runs become concurrent, long-running, and tool-intensive, this chat-centered baseline exposes projects to lost intent, cross-runtime discontinuity, ambiguous ownership, duplicated inference, workspace and integration failure, and unverifiable completion claims.

This paper frames agentic software engineering as a distributed execution problem and asks whether a human can type one complex outcome, start the agent, and walk away. Intent is the human or team’s desired outcome, motivation, constraints, priorities, unresolved questions, and completion standard, progressively represented through governed artifacts. Custody bounds task ownership and workspace authority; Trajectory records selected governance activity; and Proof provides machine-checkable validation and provenance evidence. The primary comparison remains conventional versus Decapod-governed execution under the identical natural delegation prompt.

The expanded program defines fleet coherence : independently launched agents use the same durable project authority, task custody, selected trajectory, and proof boundary across sessions, handoffs, and concurrent work. Agents and conversations are ephemeral execution processes; the governed repository is the durable coordination subject. This is not a fifth primitive, global consensus, lossless conversation transfer, or a conflict-free merge guarantee. The human speaks naturally to the available agent; the agent handles the stable governance protocol. Decapod is offered as a candidate standard agent-callable component that coding-agent harnesses could pre-bundle or discover, not as an already adopted standard or shipped vendor integration. No controlled result is reported here.

1. Intent

Persistent Goals

→

2. Custody

Isolated Boundaries

→

3. Trajectory

Governance Record

→

4. Proof

Validation Evidence

Core Primitives

Intent

Human/Team Intent

The human or team’s desired outcome, motivation, constraints, priorities, unknowns, stop conditions, and completion standard, progressively resolved through governed plans, todos, specs, and proof expectations.

Custody

Governed Custody

Task claims, task-scoped git worktrees, and configured containers reduce workspace collision and unsafe repository mutation risk; credential isolation depends on the runtime boundary.

Trajectory

Auditable Trajectory

Event-backed records for selected task, broker, proof, memory, and validation activity. Designed to support reconstruction without claiming complete runtime capture.

Proof

Verifiable Proof

Machine-checkable validation, proof-run, workunit, and provenance evidence linked to governed task state.

Fleet Coherence Research Program

Implementation basis: Decapod 207e9a3 (v0.66.3). The audit verifies atomic exclusive claims, trust-gated shared ownership and handoff, task workspaces, deterministic context bundles, local project memory, selected governance events, proof/workunit state, and publication gates. Handoff does not transfer hidden chat state; worktrees do not prove integration; the public cloud backend is unavailable.

Central fleet question: Can multiple independently launched agents concurrently solve similar but distinct problems in the same codebase while preserving shared authority, distinct task custody, isolated mutation, dependency awareness, and integrated proof? Similar work is not necessarily duplicate work, and separate worktrees do not resolve semantic overlap or prove the combined result.

Study A<br>Walk-away alignment<br>Primary CN versus DN comparison under the identical complex natural prompt.

Study B<br>Handoff continuity<br>Structured conventional handoff versus governed ownership/context/trajectory/proof transfer.

Study C–D<br>Concurrency and tool switching<br>Duplicate work, direct collisions, deferred integration failure, and reorientation across workbenches.

Study E<br>Observational dogfooding<br>At the inspected commit: 2,127 git commits and 326 tags from February–July 2026. Activity evidence only; no causal or review-burden claim.

A Portable Governance Contract for Agent Harnesses

Target interaction: the human speaks naturally to whichever agent is available; the agent invokes the governance protocol. Decapod is offered as a candidate agent-governance interoperability profile so Claude, Codex, Antigravity, Grok, and other conforming coding-agent harnesses could pre-bundle...