AI agents: clarify, act, remember, and learn across long tasksAI agents differ from ordinary model calls because their outputs change what happens next. They gather requirements, choose tools, modify environments, preserve state, and continue over many steps. That creates failure modes hidden by a final score: an unstated assumption can corrupt a plan, an early mistake can propagate, and deleted information can survive in derived state.This guide is organized by engineering problem rather than agent label. Start with clarification when an agent converts ambiguous requests into structured artifacts; use the construction and evaluation section for complete systems; read the memory section for repeated or stateful work; and use the learning section when trajectories must become training data or step-level credit.
Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive OptimizationStart here for a concrete account of how a persistent gap ledger separates missing-requirement detection from question selection, with ablations, interaction costs, and open-versus-choice results. It does not test live-user tolerance or contradictory requirements.OR-Clarify uses 100 optimization cases, hidden fact slots, bounded interaction, and a simulated user; InterOPT’s clearest win is in the choice protocol.
What are you trying to solve?
Clarify missing requirements before committing to a planAn agent can produce a valid-looking optimization model while silently inventing objectives, constraints, or business rules. InterOPT separates finding formulation-critical gaps from choosing the next question: a persistent ledger tracks what remains unresolved, while a second stage asks or stops. Completeness becomes explicit state rather than a fresh intuition on every turn.
Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive OptimizationStart here for a concrete account of how a persistent gap ledger separates missing-requirement detection from question selection, with ablations, interaction costs, and open-versus-choice results. It does not test live-user tolerance or contradictory requirements.OR-Clarify uses 100 optimization cases, hidden fact slots, bounded interaction, and a simulated user; InterOPT’s clearest win is in the choice protocol.
How can an agent decide whether to ask another question or start modeling?Track unresolved facts that would change decision variables, objectives, or constraints, and bind each question to one tracked gap. In OR-Clarify’s choice protocol, this reduced silent assumptions and improved exact recovery; premature stopping remained a larger problem than question wording alone.Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization
What trade-off comes with more systematic clarification?Coverage requires interaction. InterOPT averaged 10.7 turns in the choice setting, roughly three times the baselines. Prioritize high-severity gaps and measure both silent assumptions and question burden rather than optimizing completeness alone.Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization
Test the whole agent system under deployment constraintsAgent reliability depends on more than whether a model solves isolated tasks. A developer must interpret records, elicit requirements, work through production APIs and inherited code, and respect serving limits. Evaluation can itself become costly, so intermediate behavior may predict outcomes and stop runs early—provided results remain calibrated per agent.
$τ^τ$-Bench: An Environment for End-To-End, Realistic Agent ConstructionRead this to see agent construction evaluated as a complete client engagement rather than an isolated coding task. It identifies failures under records, API, codebase, model, and cost constraints, but does not measure real customer deployments.Results cover 53 simulated customer-service construction tasks across four domains and held-out simulated users.
EarlyEval: Cheaper Agent Evaluation via Early Outcome PredictionUse this when benchmark execution cost limits iteration. It shows how intermediate run features can trigger calibrated early stopping and reports cost and score changes; it does not show that the classifiers transfer unchanged to new benchmarks.Evaluated on SWE-bench Verified, TerminalBench, and Toolathlon with benchmark-specific intermediate features.
What does an end-to-end construction benchmark reveal that task benchmarks miss?It exposes failures in requirements communication, record comprehension, architecture exploration, and serving-cost choices. In τ^τ-bench, the strongest reported configuration passed 23.9% of evaluation simulations versus an 82.2% expert reference ceiling across 53 customer-service construction tasks.$τ^τ$-Bench: An Environment for End-To-End, Realistic Agent Construction
How can teams reduce execution cost within retained benchmark tasks?EarlyEval trains success and failure classifiers on intermediate behavioral, textual, and reference-solution features, then halts when calibrated confidence crosses a threshold. Across three benchmarks it removed 13–26% of steps while average per-agent resolve rates moved by one to two percentage points.EarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction
Engineer memory for reuse, migration, and revocationAgent memory has distinct jobs. A skill library preserves procedures across GUI tasks; a durable knowledge store must survive model and embedding changes; and a forget request must remove downstream influence from summaries, plans, and caches. Treating all three as text retrieval obscures their different mechanisms and tests.
From Interaction Traces to Persistent Skills: Online Evolution for Computer-Use AgentsRead this for an auditable way to convert GUI traces and feedback into persistent, versioned procedures. It exposes cross-task retrieval and revision churn, making it useful for studying both benefits and maintenance failures.The comparison uses four OSWorld application domains, a fixed action-generation and GUI-grounding stack, and a five-iteration warm-up.
Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory PortabilityUse this to plan memory migration tests across model writers and embedding versions. It separates construction loss from retrieval failure and tests repair with and without source histories; it does not cover larger models or natural user histories.Controlled evidence uses 48 synthetic histories, exact randomized codes, and two open-weight models below 10B parameters.
Forgetting Without Restarting: Execution-State Unlearning for Stateful LLM AgentsRead this when deletion must remove behavioral influence, not only stored plaintext. It formalizes which execution suffix must be recomputed and evaluates provenance-guided replay; it does not establish legal compliance or nondeterministic production guarantees.The formal lower bound assumes deterministic transitions; audits span three agent suites, nine baselines, and three model families.
How can a computer-use agent reuse procedures without changing model weights?Convert trajectories and evaluator feedback into a versioned skill library, freeze a snapshot for each iteration, and expose accepted updates later. Across four OSWorld domain runs, the evolving library beat a matched empty-library control after warm-up, but GIMP revisions showed that repeated accepted edits need not restore the task that produced a skill.From Interaction Traces to Persistent Skills: Online Evolution for Computer-Use Agents
What makes an agent memory store portable across model upgrades?Prefer explicit structure, isolate embedding versions, test each migration direction, and retain raw source history for repair. In the controlled study, fixed-schema graphs changed little after a writer swap, while compressed notes were direction-sensitive and a half-migrated RAG index captured less than half the gain of full re-embedding.Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory Portability
Why can an agent still act on information after its memory record is deleted?The information may already have influenced summaries, pending plans, prompts, and the KV cache. Provenance-guided selective replay restores a checkpoint before the information arrived and regenerates later state from sanitized inputs, matching full-reset behavior in the reported audits with fewer recomputed tokens.Forgetting Without Restarting: Execution-State Unlearning for Stateful LLM Agents
Turn long trajectories into targeted learning signalsA completed run can support learning at different resolutions. DRACO turns multi-criteria trajectory judgments into differentiated per-step advantages. NeoHorse-1 records routing decisions, interactions, reasoning, and tool calls, filters and labels them, then uses evaluation feedback to choose the next training mixture. These are separate mechanisms for credit assignment and iterative data selection.
DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent TrainingUse this for long-horizon policy training when completed trajectories can be judged by rubrics but no programmatic verifier is available. It explains closed-form redistribution from criterion scores to per-step advantages.Evidence covers GRPO training on AppWorld and out-of-domain evaluation on Tau-Bench.
NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing HarnessRead this for a concrete evaluation-selection-update loop: routing traces are validated and labeled, organized into curricula, and used to allocate later post-training data. It demonstrates an initial feedback-driven loop rather than unrestricted recursive self-improvement.Reported evidence spans eleven agent, tool-use, coding, and instruction-following benchmarks with 4B and 9B models.
How can trajectory-level judgments become step-level training signals?Annotate completed trajectories with rubric criteria, associate criteria with responsible steps, and redistribute the judgments rather than applying one scalar uniformly. DRACO uses closed-form redistribution in GRPO and requires no separately trained attribution model.DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training
How can evaluation feedback determine what an agent learns from next?NeoHorse-1 records predicted capability demand, routing choices, and ensuing interactions; validates and labels those traces; organizes them into a curriculum; and reallocates the next training mixture from evaluation results. Reported 4B and 9B gains show an initial closed feedback loop, not unrestricted autonomous self-improvement.NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness