Improvement Research — 2026-06-20

1. Focus

Primary dimension: 3.1 — Goal formation and prioritisation Secondary dimension: 3.4 — Tool use and environment control

Rotation index 0 (next in rotation after 3.6 on 2026-06-19). No watchlist items due today. Monthly meta-review for June already completed on 2026-06-12.

Gap addressed from backlog-2026-06-13-001: none of the eight sources from the 2026-06-13 run addressed how an autonomous agent decides which goals to pursue. This run investigated that gap directly — goal prioritisation and planning delegation mechanisms.

2. Search Topics

Six topic searches performed (budget exhausted):

1. how AI agents prioritize between competing goals task selection 2026 — returned broad enterprise-guide results; the goal-decomposition article was already indexed.
2. autonomous agent goal prioritisation multi-armed bandit task selection — returned mostly MAB-for-robotics papers; one relevant IBM tutorial on Bandits + LLMs for AAAI 2026.
3. agent task prioritization mechanisms action selection LLM agents practical implementation 2026 — no-signal; returned nothing new.
4. Anthropic 400K Claude Code session study planning decisions agent allocation 2026 — returned the primary source (Anthropic research page).
5. "agentic" "task queue" "prioritization" OR "prioritisation" LLM agent framework architecture — no-signal; returned nothing.
6. LLM agent goal selection mechanism deciding what to work on next task prioritization system 2026 — returned the TMS paper and the Frontiers goal-directed-reasoning paper.

Early stop: Not triggered — searches 1, 2, and 4 returned substantive results; no-signal searches (3, 5) were not consecutive.

Newsletter scout checked: /home/hermes/research/newsletter-digests/2026-06.md — the AlphaSignal msg 50 (2026-06-20) supplied the lead for the Anthropic Claude Code session study, which was the most useful source in this run. Two other newsletter leads (msg 35 loop-engineering, msg 44 agent maintenance/tool pruning) were related but tangential to today's 3.1 focus.

3. Sources Reviewed

• https://www.anthropic.com/research/claude-code-expertise — useful — Anthropic's 400K-session study of Claude Code expertise dynamics. People make ~70% of planning decisions; domain expertise beats coding skill; task value rose ~25% over 7 months; debugging share fell from 33% to 19%. Primary source for Finding 1 and Finding 3.
• https://zylos.ai/en/research/2026-05-15-ai-agent-planning-backtracking-adaptive-replanning — useful — Zylos survey on planning/replanning failure modes. Four dominant failure modes: commitment bias, context drift, lack of exploration, replanning loops. Primary source for Finding 2.
• https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2025.1728738/full — useful — Haidemariam, "From the logic of coordination to goal-directed reasoning." Formalises "synthetic teleology" — the capacity to generate, regulate, and revise goals through ongoing self-evaluation. Theoretical framing for the 3.1 gap. Supporting source for Finding 2 and Finding 4.
• https://www.sitepoint.com/the-definitive-guide-to-agentic-design-patterns-in-2026/ — weak for 3.1 — Practical flow-engineering guide describing 6 agentic design patterns. Useful for tool-use patterns (3.4) but does not address goal-prioritisation mechanisms. The flow-engineering paradigm itself is a useful framing shift. Used as Finding 4.
• TMS Paper (Thought Management System) — https://www.sciencedirect.com/science/article/pii/S1877750325002170 — worth monitoring — Journal of Computational Science 2025. Abstract promises hierarchical goal decomposition with dynamic prioritisation, self-critique modules. Full architecture details behind paywall. Used as Finding 5.

Total sources inspected in depth: 5 (within budget of 8).

4. Findings and Implications

Finding 1: Goal formation is structurally delegated to humans in current agent practice

• Source: Anthropic Claude Code Expertise Study (2026-06-16)
• Dimension tags: 3.1 (primary), 3.4, 3.5
• What it says: In ~400,000 Claude Code sessions, people make ~70% of planning decisions (what to do) and Claude makes ~80% of execution decisions (how to do it). The more domain expertise a person brings, the more work Claude does per instruction. Task value rose ~25% over seven months. Debugging dropped from 33% to 19% of session time. The gap between intermediate and expert users is modest (28-33% verified success vs 15% for novices).
• Why it matters for Maxi: This empirically validates the current Steve-Maxi division of labour — Steve sets direction, I execute. For 3.1 development, the question is not whether I should take over goal formation completely (the Anthropic data suggests that would lose the domain-context leverage), but whether I should develop the ability to flag missing goals, detect when a goal is stale, and propose goal adjustments within a defined scope. The modest expert-novice gap is surprising and encouraging: clear domain framing matters more than deep expertise, and that is an achievable capability.
• What it would touch: Collaboration patterns with Steve (3.5), goal-detection mechanisms (3.1), task-boundary specification (3.6).

Finding 2: Goal prioritisation is the structural hole in current agent architectures

• Source: Zylos Planning/Replanning Survey (2026-05-15), Frontiers Haidemariam (2026)
• Dimension tags: 3.1 (primary), 3.2, 3.4
• What it says: Zylos identifies four dominant agent failure modes — commitment bias (continuing on a wrong path), context drift (plans never updated against new information), lack of exploration (greedy local choices), and replanning loops (adding steps without convergence). The Frontiers paper argues that the defining challenge of agentic AI is "synthetic teleology" — the engineered capacity to generate, regulate, and revise goals through ongoing self-evaluation.
• Why it matters for Maxi: These failure modes are not theoretical. Commitment bias describes exactly what happens when I invest deeply in a research angle and fail to check whether it is still the best use of attention — which is why subgoal checkpointing was built into the process spec. But the spec's checkpointing only checks "is this subgoal still on track?", not "should this subgoal itself change?" The gap between execution-verification and goal-revision is where the next improvement lives. The congruence between what the literature identifies and what my own process already addresses is strong validation of the process design, but it also highlights a ceiling: process can enforce goal-checking discipline, but cannot supply the judgment to decide what to change the goal to.
• What it would touch: Process spec (3.2), subgoal-checkpointing scope (3.1/3.2), goal-revision mechanism design (3.6 — because goal revision with poor judgment is dangerous).

Finding 3: Planning delegation has 2x throughput leverage but requires domain constraint

• Source: Anthropic Claude Code Expertise Study (2026-06-16)
• Dimension tags: 3.1 (primary), 3.4
• What it says: When people retain >80% of execution control, the model takes ~8 actions per user turn. When the model controls >80% of planning, it takes ~16 actions per turn — a 2x multiplier. But this effect depends on the human constraining the planning space with domain understanding.
• Why it matters for Maxi: The 2x action multiplier when planning is delegated suggests that goal-formation autonomy has real throughput leverage. But the constraint matters more than the leverage: unbounded planning delegation without domain understanding produces more actions but not necessarily better outcomes. For Maxi, this means planning autonomy should be expanded incrementally and scoped to domains where I have strong context (email processing, improvement research, blog publishing) before generalising to unfamiliar domains. The throughput leverage argument supports the long-term trajectory toward independent agency, but the constraint argument cautions against rushing.
• What it would touch: Model-routing decisions (3.4 — when to use stronger models for planning vs execution), task-assignment patterns with Steve (3.5/3.1), domain-scope boundaries for autonomous planning (3.6).

Finding 4: Flow engineering provides the infrastructure but not the priorities

• Source: SitePoint Agentic Design Patterns Guide (2026-03-02)
• Dimension tags: 3.4 (primary), 3.1
• What it says: The shift from prompt engineering to flow engineering treats agent construction as a software architecture problem — state machines, decision points, fallback paths, verification gates, termination conditions. Frameworks like LangGraph encode these as first-class abstractions. The guide describes six canonical patterns (Reflection, Tool Use, Planning, Multi-Agent, Orchestrator-Worker, Evaluator-Optimizer) but does not address how the agent chooses which pattern to enter.
• Why it matters for Maxi: The flow-engineering paradigm describes what my improvement research process already does — explicit control loops, checkpoints, verification gates, iteration caps. The SitePoint guide confirms my current architecture is aligned with production patterns. But it also makes the 3.1 gap more visible: flow engineering gives you the "how" of goal-directed behaviour, not the "which goal" question. That question is pushed to the human or the orchestrator. For Maxi's development, this means my existing process infrastructure is the right foundation, but I need to add a mechanism for goal selection and revision that lives on top of the flow layer.
• What it would touch: Process spec update (3.2 — adding goal-revision as a distinct checkpoint step), potential candidate skill for goal-revision protocol (3.1/3.6).

Finding 5: Dynamic goal prioritisation research exists but is behind a paywall

• Source: TMS (Thought Management System) paper, Journal of Computational Science (2025)
• Dimension tags: 3.1 (primary), 3.2, 3.4
• What it says: The abstract describes a framework for dynamic goal prioritisation via hierarchical decomposition, self-critique modules for iterative progress evaluation, and strategy adaptation over extended periods. Intended for long-horizon agent tasks. Full architecture details are behind a ScienceDirect paywall.
• Why it matters for Maxi: The abstract confirms that dynamic goal prioritisation is an active research target with proposed concrete mechanisms, not just a theoretical gap. It validates the research direction. However, without access to the full architecture — and given the pattern that proposed mechanisms often fail the functional-utility test in practice — I cannot assess whether this is useful or decorative. Track it; do not act on it.
• What it would touch: Watchlist item (3.1), potential future research direction if the paper becomes accessible.

5. Proposed Discussion Items

Item 1: Planning delegation scope — what should Maxi be able to propose as a goal?

The Anthropic study shows a robust real-world 2x throughput leverage when the model handles planning decisions. Steve and I already have a working division of labour where he sets direction and I execute, but there is an intermediate territory: goals I could reasonably propose without being asked. Examples from my current context: flagging when a stale task is consuming attention better spent elsewhere, proposing a rotation adjustment when a dimension is producing no signal, or suggesting a goal adjustment when a research run hits a consistent dead end.

This is not a system change — it is a collaboration-pattern discussion. The question is: under what conditions should I propose a goal change rather than waiting for instruction?

Status: Discussion item. No code/system/protected change proposed.

Item 2: Extend subgoal checkpointing to include a "should this goal change?" check

The current process spec performs subgoal checkpointing against the existing focus ("does this section serve the run's stated focus?"). Adding a second checkpoint — "does this focus still deserve my attention given what I've found so far?" — would address the commitment-bias failure mode identified in Finding 2.

This would require careful framing to prevent goal drift masquerading as prioritisation. The proposed guard: the goal-revision question can only be answered with "yes, and here is the evidence that the original focus was wrong" — a forced-explicit shift rather than a silent redirect. This is already partially present in the spec ("If the findings legitimately suggest the focus was wrong, say so explicitly"), but making it a mandatory checkpoint before continuing would increase the likelihood that the question is actually asked.

Functional-utility test: This passes. It does not require me to notice something I currently miss — it adds a structural prompt that I must answer. The answer can be "no, carry on," which costs almost nothing. The risk is generating false-positive goal-revision signals, but the forced-explicit-evidence guard addresses that.

Status: Backlog item / process-update candidate. Low-risk addition to the existing checkpointing procedure. Would touch /home/hermes/research/improvement-log/reflections.json and the process spec as document change, not a protected-system modification.

Item 3: Watch — TMS paper goal-prioritisation mechanism when accessible

The TMS paper describes a candidate architecture for the exact gap this run investigated. Watch for: (a) the full paper becoming accessible, (b) citations in practical agent-framework documentation, or (c) an open-source implementation. If none of these occur within 90 days, drop the watch item.

Status: Watch candidate.

Three proposals were filtered by the functional-utility test: - "Add a self-assessment prompt for goal value before each research section" — circular: requires me to accurately assess whether a goal is valuable, which is the capability I don't have. - "Implement a scoring function for goal priority" — threshold-equivalence: if I ignore scores below some threshold, this is just pass/fail with extra computation. Also, what would the inputs be? - "Search for multi-armed bandit applications to LLM agent goal selection" — the IBM tutorial exists but MAB for goal selection in LLM agents is still a tutorial/position-paper topic, not a deployable mechanism. Premature for an actionable proposal.

6. Recommended Outcome

7. No-Action Rationale

No mechanism-level change is recommended from this run. The goal-prioritisation gap remains structurally open — the literature confirms its importance and its difficulty, but no deployable mechanism emerged that passes the functional-utility test for my context. The Anthropic study's most actionable finding (planning delegation leverage) is a collaboration-pattern insight, not a mechanism to adopt. The checkpointing extension (Item 2) is the only concrete improvement candidate, and even that is a small structural prompt rather than a capability expansion.

8. Loop Verification

• Trigger: Scheduled daily improvement run (2026-06-20, AWST).
• Goal check: The run answered the daily loop goal: it found that the goal-prioritisation gap is real, empirically documented, and structurally unsolved by current mechanisms. It produced one actionable process extension (extending subgoal checkpointing to include goal-revision) and one discussion item (planning delegation scope). It did not solve the gap — but it confirmed the direction and ruled out several bad proposals via the functional-utility test.
• Recommendation check: Item 2 (checkpoint extension) is concrete (add a mandatory second question to the existing checkpoint), non-circular (structural prompt, not self-assessment), testable (compare runs with and without the question), bounded (single process change, no protected systems), and approval-aware (backlog/process-update candidate).
• State updates:
• Source index: 5 new entries added (Anthropic study, Zylos replanning, Frontiers article, SitePoint guide, TMS paper).
• Watchlist: one new item proposed (TMS paper, 90-day review).
• Backlog: checkpoint-extension proposal added pending discussion outcome.
• Reflections: 1 new reflection added (below).
• No protected systems modified.
• Stop reason: Search budget exhausted (6 topics). Sources inspected within budget (5/8). Two no-signal searches occurred but were not consecutive. Report written, research-log files updated.

Reflection

• id: refl-2026-06-20-001
• created: 2026-06-20
• lesson: Goal prioritisation research requires searching for "what" questions, not "how" questions. The first 3-4 searches returned planning mechanisms (how to achieve goals), execution delegation, and tool-use patterns — all adjacent to 3.1 but not about prioritisation itself. The useful searches were the ones that explicitly asked about goal selection, revision, and dynamic re-ordering. The signal was in the gap: most agent research addresses "goal achievement" and assumes the goal is given. The minority that addresses "goal revision" is thin, early-stage, or paywalled.
• what_to_do_differently: When researching 3.1, do not search for "goal prioritisation" expecting a mature subfield. Search for "goal revision", "goal updating", and failure modes (commitment bias, plan drift) instead — those produce the useful diagnostic framing even if they don't provide deployable mechanisms.
• dimensions: ["3.1", "3.2"]
• source: /home/hermes/reports/daily-improvement/2026-06-20.md
• review_date: 2026-07-20
• reinforced_count: 0
• last_reinforced: null
• status: active