product docs · source date 2026-07-14 · 0 comments ·
original1. Problems / challenges / motivations
- Optimization research assumes someone will wire traces, evals, candidate generation, and deployment together by hand.
- Enterprises will not run a self-improvement loop without lineage, diffs, and rollback.
- Production traces are the best optimization signal available and usually go unused.
2. Key ideas
- A managed...
researcher blog · source date 2026-07-04 · 0 comments ·
original1. Problems / challenges / motivations
- Discussion of recursive self-improvement defaults to model weights, while most of the practical gains in 2026 came from the code around the model.
- The relevant literature is scattered across context engineering, scaffold search, and Gödel-machine work with no common map.
- Practitioners have no shared vocabulary...
arXiv paper · source date 2026-06-24 · 0 comments ·
original1. Problems / challenges / motivations
- Self-improving agents are bounded by the evaluator that scores their modifications, so a fixed evaluator caps improvement.
- In hard-to-verify domains — paper writing, proof grading — there is no benchmark to serve as the evaluator at all.
- Letting the evaluator evolve freely invites the agent and evaluator to drift...
engineering blog · source date 2026-06-03 · 0 comments ·
original1. Problems / challenges / motivations
- Self-service analytics fails in two directions: give an agent broad data access and metric definitions diverge across teams; lock access down and it cannot answer the business questions people actually have.
- Three specific failure modes drive wrong answers: concept-to-entity ambiguity (mapping a question to the...
peer-reviewed paper · source date 2026-06-01 · 0 comments ·
original1. Problems / challenges / motivations
- Best-of-N spends the same verification compute on easy and hard candidates alike.
- Generative verifiers are accurate but expensive, which limits how often they can be called.
- Verifier training and verifier-guided search are usually studied separately.
2. Key ideas
- FlexiVe is a generative verifier that...
open-source framework · source date 2026-06-01 · 0 comments ·
original1. Problems / challenges / motivations
- Every lab reimplementing PPO/GRPO plumbing wastes effort and makes results incomparable.
- Multi-turn agentic rollouts with tool calls do not fit training loops designed for single-turn completions.
- Rollout throughput becomes the bottleneck without async and load balancing.
2. Key ideas
- A production RL training...
peer-reviewed paper · source date 2026-06-01 · 0 comments ·
original1. Problems / challenges / motivations
- Production agent quality is gated by the quality of the evaluation loop, and most teams' LLM judges are not calibrated against anything.
- Hand-written judge prompts drift and disagree with human reviewers, which makes every downstream decision noisy.
- Optimization research is overwhelmingly benchmark-based, with...
engineering blog · source date 2026-05-27 · 0 comments ·
original1. Problems / challenges / motivations
- Vertical agents fail in domain-specific ways that generic benchmarks never surface.
- Practitioner corrections are the highest-value signal a deployed agent produces and usually vanish into a support queue.
- "Self-improving" typically means prompt tweaks, which does not fix root causes in the surrounding system.
2....
product release · source date 2026-05-19 · 0 comments ·
original1. Problems / challenges / motivations
- IDE-centric tooling assumes a human in the editor, which is the wrong center of gravity once agents do most of the editing.
- Steering long-running agents needs surfaces for both synchronous and asynchronous work.
- Model and harness are usually developed by different teams and integrated late.
2. Key ideas
- A...
arXiv paper · source date 2026-05-01 · 0 comments ·
original1. Problems / challenges / motivations
- Retrieval-style memory injects a blob of past experience ahead of time, most of which is irrelevant to the current step.
- Irrelevant memory costs tokens, adds latency, and actively distracts the policy.
- Vector recall quality is not the same thing as memory being useful to an acting agent.
2. Key ideas
- Replaces...
product docs · source date 2026-05-01 · 0 comments ·
original1. Problems / challenges / motivations
- Prompt optimizers mostly live inside one framework, so adopting them means adopting the framework.
- Optimized prompts are artifacts that need versioning, diffing, and rollback like any other deployed asset.
- Without CI integration, optimization is a one-off experiment rather than a maintained process.
2. Key...
peer-reviewed paper · source date 2026-05-01 · 0 comments ·
original1. Problems / challenges / motivations
- Multi-agent debate and collaboration results were mostly established on weaker base models and assumed to hold as models improve.
- Teams have no decision rule for when to add agents versus when a single strong model suffices.
- Collaboration adds cost and latency that need justification.
2. Key ideas
- Quantifies...
peer-reviewed paper · source date 2026-04-24 · 0 comments ·
original1. Problems / challenges / motivations
- RL post-training for compound LLM systems is rollout-hungry: it learns from a scalar reward and throws away everything the trace actually said about the failure.
- Prompt optimization by hand does not scale across a multi-module pipeline.
- Earlier automatic optimizers (MIPROv2) search over instructions without...
open-source tool · source date 2026-04-01 · 0 comments ·
original1. Problems / challenges / motivations
- RL on agent tasks needs a reward function, and most real tasks have no verifier and no labeled data.
- Hand-crafting reward functions per task is the main cost of applying RL to agents.
- Absolute LLM-judge scores are noisy and poorly calibrated across tasks.
2. Key ideas
- A general-purpose reward function that...
research blog · source date 2026-04-01 · 0 comments ·
original1. Problems / challenges / motivations
- Multi-agent systems are adopted on the assumption that more agents means better results, with little controlled evidence.
- Comparisons across papers confound topology, model, and token budget.
- Errors propagate differently depending on how agents are wired together, and that effect is rarely measured.
2. Key...
open-source framework · source date 2026-03-29 · 0 comments ·
original1. Problems / challenges / motivations
- RL post-training couples a training engine to a rollout engine, and synchronous coupling leaves accelerators idle.
- Agentic rollouts (tool calls, sandboxes, multi-agent loops) are slow and variable, which makes the coupling worse.
- Agentic RL is often bolted on as a separate framework rather than treated as data...
practitioner blog · source date 2026-03-03 · 0 comments ·
original1. Problems / challenges / motivations
- Eval methodology is well documented and badly practiced, because doing it properly is tedious.
- Coding agents are already in the loop but are not taught how to build evaluation systems.
- Error analysis — the highest-value step — is the one teams skip first.
2. Key ideas
- Publishes `evals-skills`, a plug-in set of...
engineering blog · source date 2026-02-11 · 0 comments ·
original1. Problems / challenges / motivations
- Codebases are written to be legible to humans, which is not the same as being legible to a coding agent.
- When agents write most of the code, the bottleneck moves from writing to specifying, validating, and observing.
- Agent failures in a large repo are hard to diagnose without per-run observability.
2. Key...
model release · source date 2026-02-05 · 0 comments ·
original1. Problems / challenges / motivations
- Long-horizon agent work needs both strong coding and strong reasoning, which had been split across separate models.
- Fire-and-forget agent runs waste time when the user can see it going wrong but cannot intervene.
- Prompting guidance for agent models is under-specified relative to how much it changes behavior.
2....
technical report · source date 2026-02-01 · 0 comments ·
original1. Problems / challenges / motivations
- RLVR works where rewards are verifiable, which excludes most multimodal and open-ended agentic tasks.
- Scaling agentic post-training needs a reward signal that does not depend on hand-written verifiers for every domain.
- Single-agent rollouts limit the diversity of trajectories available to learn from.
2. Key...
peer-reviewed paper · source date 2026-01-30 · 0 comments ·
original1. Problems / challenges / motivations
- Rewriting a whole system prompt on each optimization step introduces brevity bias: the rewrite compresses away details it does not currently see a use for.
- Repeated rewrites cause context collapse, where accumulated knowledge degrades rather than compounds.
- Adapting a system to a new domain by retraining is slow...
engineering blog · source date 2026-01-22 · 0 comments ·
original1. Problems / challenges / motivations
- Building a specialized agent per domain multiplies maintenance and fragments capability.
- Stuffing all domain knowledge into a system prompt does not scale past a few domains.
- Reusable agent capability had no portable packaging format.
2. Key ideas
- Explains why Anthropic stopped building specialized agents and...
engineering blog · source date 2025-11-04 · 0 comments ·
original1. Problems / challenges / motivations
- Loading every MCP tool definition into context burns tokens before the agent does anything.
- Intermediate tool results flow through the model even when the model only needs a filtered summary.
- Tool selection degrades as the tool count grows.
2. Key ideas
- Presents MCP servers as code APIs the agent explores on a...
peer-reviewed paper · source date 2025-10-24 · 0 comments ·
original1. Problems / challenges / motivations
- Self-modification search scores a variant by its own benchmark performance, but the useful property is whether it leads somewhere — the two come apart.
- The paper names this the Metaproductivity–Performance Mismatch: a mediocre variant can be the ancestor of the best ones.
- Expanding the wrong branches burns large...
engineering blog · source date 2025-09-29 · 0 comments ·
original1. Problems / challenges / motivations
- Prompt engineering assumes a fixed instruction; agents assemble a different context at every step.
- Long-horizon agents run past any context window, so something has to decide what survives.
- The field lacked shared vocabulary for what that decision process is.
2. Key ideas
- Frames context engineering as the...
peer-reviewed survey · source date 2025-09-02 · 0 comments ·
original1. Problems / challenges / motivations
- Agentic RL work is spread across planning, tool use, memory, reasoning, and perception with no shared taxonomy.
- Standard LLM RL formulations assume a single-step MDP, which does not describe an agent acting over a long horizon with partial observability.
- Environments, benchmarks, and frameworks are fragmented and...
technical report · source date 2025-07-01 · 0 comments ·
original1. Problems / challenges / motivations
- Long context windows are marketed as capacity, and teams treat "fits in the window" as "will be used correctly."
- Needle-in-a-haystack tests use high semantic similarity between question and needle, which is the easy case.
- Without evidence of degradation, there is no case for the engineering cost of compaction and...
peer-reviewed paper · source date 2025-05-30 · 0 comments ·
original1. Problems / challenges / motivations
- The classical Gödel machine requires a formal proof that a self-modification is an improvement, which is infeasible for real systems.
- Agent design is currently done by hand, one scaffold change at a time.
- Greedy self-modification collapses into a local optimum and loses variants that would have paid off...