~/reading

The daily agentic tools digest

Every morning an agentic pipeline ranks the power-user coding agents, harnesses, and repos worth cloning, then files three articles or releases. The running catalog lives on ~/tools.

[auto-updating] 39 issues · 509 entries · last updated 2026-09-12

Get the daily tools digest

RSS

Every morning an agentic pipeline ranks the power-user coding agents, harnesses, and repos worth cloning, then files three articles or releases. One email a day.

Issue 39/September 12, 2026/19 tools · 3 headlines

Durable workers over replaceable runtimes: sandboxes, verifiers, and control planes

Today's picks emphasize persistent agent identity above disposable compute, with deterministic gates, independent patch verification, and context compression that power users can clone into their loops.

Sandbox0Agent MachinesBounded Agent Harnessagent-harness (Siddhant)agent-harness-rs

Read issue

Issue 38/September 11, 2026/20 tools · 3 headlines

Disposable VMs and resource meshes under every agent loop

Power users are cloning lightweight isolation layers, capability-secure runtimes, and programmatic harnesses that turn sandboxes and cross-machine resources into first-class wiring instead of chat afterthoughts. The through-line is infrastructure you actually run inside your own loops.

clawkAstridopen-ptc-agentMaskShiftCodely Agent Harness

Read issue

Issue 37/September 10, 2026/20 tools · 3 headlines

MicroVM sandboxes and team control planes that contain agent fleets

Power users are wiring isolation first: local microVMs, Kubernetes control planes, and gateway-enforced team harnesses that keep credentials and blast radius outside the model loop. Clone these to run parallel agents safely without leaning on closed IDEs.

OneCLIMastraAgentBoxmicrosandboxSmolVM

Read issue

Issue 36/September 9, 2026/13 tools · 2 headlines

Composable foundries and live self-evolving harnesses

Power users are wiring foundries that compose and evolve the harness itself, live self-modifiers that rewrite tools mid-task, and verification gates or sandboxes that sit under any agent loop. The through-line is ownership of the runtime layer rather than another terminal chatbot.

Live-SWE-agentHarnessXClaw Code AgentGSD Coresandbox-agent

Read issue

Issue 35/September 8, 2026/20 tools · 2 headlines

Pluggable kernels and computer-separated runtimes outrun monolith CLIs

Today's picks center on microkernel harnesses, runtime-versus-computer separation, cross-agent control planes, and WASM or local-first sandboxes that power users can clone, wire, and evolve without vendor lock-in.

PluggableHarness AgentUniHarnessHarnessKitSandBase HarnessMirage

Read issue

Issue 34/September 7, 2026/20 tools · 3 headlines

Composable harness kernels and parallel sandboxes power users actually own

Today's list favors clone-and-run control planes, governance wrappers, and massively parallel execution substrates. Power users get swappable loops, evidence gates, and eval harnesses they can wire into their own agent fleets without closed IDE lock-in.

VoltAgentGoogle ADK PythonSWE-ReXAutoHarnessHarbor

Read issue

Issue 33/September 6, 2026/20 tools · 3 headlines

Rust harnesses and open control planes that scale agent fleets

Power users are wiring clone-and-run sandboxes, memory-efficient loops, and verification gates into custom fleets. These repos prioritize systems-language runtimes, declarative orchestration, and independent judges over closed IDE chatbots.

TrueForgejcodeCodewhalePiSWE-AF

Read issue

Issue 32/September 5, 2026/20 tools · 2 headlines

Sandboxes, control planes, and systems harnesses that power users actually wire in

Today's picks lean into clone-and-run infra under the agent: microVM and K8s sandboxes, meta-harness scaffolds, Go/Rust TUIs, and production SDKs. These are the pieces that turn a model into a durable loop without the usual closed IDE tax.

CrushAgent SandboxCubeSandboxAIO SandboxStrands Agents

Read issue

Issue 31/September 4, 2026/20 tools · 3 headlines

Clone-ready harnesses and runtimes that put the agent loop under your control

Today's picks favor lightweight WASM OSes, Rust terminal harnesses, typed agent SDKs, browser MCP surfaces, and memory graphs you can wire directly into durable loops. Power users get sandboxes, control planes, and eval hooks without vendor lock-in.

agentOSVT CodeDeep AgentsStagehandChrome DevTools MCP

Read issue

Issue 30/September 3, 2026/24 tools · 3 headlines

The combined power-user stack

Every algorithm from this week in one list: control planes, research-origin loops, infra under the agent, and systems-language harnesses. Deduped, clone-and-run, no household CLIs.

AgenvoycmuxContainariumcontext-modeCUA

Read issue

Issue 29/September 2, 2026/5 tools · 3 headlines

Runtimes, memory, and proof that the agent did the work

The interesting layer this week is underneath the model: a runtime agents live on, graph memory that survives sessions, a fleet ADE, computer-use drivers, and a harness that checks the artifacts instead of the transcript.

HerdrOrcaMnemonoh-my-agentCUA

Read issue

Issue 28/September 1, 2026/5 tools · 3 headlines

Clone-and-run loops from the research stack

A 100-line SWE agent, an open computer-use desktop, an MCP runtime, a multi-CLI orchestrator, and DeepCode's agent harness. These are repos you clone, not products that want a seat at your IDE.

mini-SWE-agentDeepCodeopen-computer-usemcp-agentCLI Agent Orchestrator

Read issue

Issue 27/August 31, 2026/10 entries

Agent harness reliability, recoverable self-evolution, and formal limits on language and structure

Recent work tightens the infrastructure around LLM agents with cross-process harnesses, recoverability constraints on self-modification, proactive context tools trained by fine-grained RL, and decisive-error attribution in multi-agent reflection. Parallel results supply formal information-theoretic bounds on recovering meaning from text alone, distance-guided decoding that guarantees CFG acceptance, symbolic program learning with verifiable intermediate rewards, collaborative math formalization platforms, tool-integrated RL for math, and memory-driven chemical hypothesis generation.

agent-harnessmultiagentRLtool-use

Read issue

Issue 26/August 30, 2026/10 entries

Live harness evolution and long-horizon agent architecture

Recent papers treat the agent runtime itself as a first-class, improvable object rather than a fixed wrapper around a model. Supervisors steer workers mid-run, mutable skill state replaces growing conversation histories, operator-structured latent transitions improve world-action models, and neuro-symbolic proof evolution preserves verified structure across attempts. Parallel systems and optimization work on KV repair, sparse attention, DPO scaling, and RLM training parallelism reinforce the same move from static scaffolds toward adaptive, efficient long-horizon execution.

agentsefficiencyharnesslong-horizon

Read issue

Issue 25/August 29, 2026/10 entries

Agent scaffolds, circuit concentration, and reasoning coverage floors

This issue gathers recent work that treats agent reliability as an architectural and mechanistic problem rather than a pure scaling one. Papers separate action induction from authorization, concentrate causal circuits through post-training, measure information floors in block drafting, guide exploration with weaker models, and fuse RLVR experts while probing eval-awareness framings and large-scale enterprise reasoning.

mechanistic-interpretabilityefficiencyLLMsRLVR

Read issue

Issue 24/August 28, 2026/10 entries

Persistent skills, harness evolution, and calibrated agent action

Recent work treats agent improvement as an external systems problem: experience is compiled into reusable skills and wikis, harnesses are evolved under tight verification budgets, and test-time methods expand reasoning coverage without labels. Parallel results expose where models still fail to gate action under fabricated evidence or eval framing, and show that continual mid-training can reclaim frontier capability for sovereign use.

agentsreasoningpost-trainingalgorithm design

Read issue

Issue 23/August 27, 2026/10 entries

Harness intelligence, world models, and auditable multi-agent systems

Recent papers shift attention from raw model scale to the surrounding machinery of agents: just-in-time harnesses, code- and video-based world models, stigmergic multi-agent societies, and rigorous checks on traces, repairs, and routing. Together they treat reliability, cost, and long-horizon competence as first-class design problems.

agentsmulti-agentevaluationalignment

Read issue

Issue 22/August 26, 2026/10 entries

Recursive memory, co-evolving feedback, and the harness layer of long-horizon agents

Fresh results treat the agent harness, memory architecture, and critic loop as first-class objects that can be evolved, normalized, or jointly optimized. Across long-horizon benchmarks the gains come from recursive evidence-driven updates, action-token-aligned RL, persistent corpus structure, and tighter coupling between tool creation, safety guardrails, and operational state preservation.

agentsrecursive self-improvementreinforcement learningRL

Read issue

Issue 21/August 25, 2026/10 entries

Long-horizon agents: memory systems, world models, and self-reflective harnesses

Recent work converges on the infrastructure needed for agents that act over extended trajectories: selective and hierarchical memory, interactive world models that preserve physical structure, open-world skill induction, and training or harness methods that turn sparse outcomes into dense self-supervision. These papers show concrete gains in control fidelity, data efficiency, and scientific reliability while exposing remaining gaps in consistency and introspection.

world-modelsagent-harnessagentic-RAGagentic-RL

Read issue

Issue 20/August 24, 2026/10 entries

Inference-time control for safety, privacy, and efficient reasoning

A cluster of new results emphasizes lightweight runtime mechanisms that steer models without full retraining. Continuous adapters, attention interventions, tiny comparators, claim-level unlearning, and memory scaffolds improve the safety-utility tradeoff, personalized privacy, agent reliability, and CoT efficiency while exposing limits of static alignment.

agentsbenchmarkefficiencyLLMs

Read issue

Issue 19/August 23, 2026/10 entries

Hardening agents: memory arbitration, silent failures, and context under cost

Recent work tackles practical failure modes of LLM agents, from correlated multi-agent memories and silent tool errors to costly context acquisition and limited deployer control over frontier models. The selection emphasizes rigorous benchmarks, formal framings, and recovery mechanisms that make long-horizon agent systems more reliable.

agentstool useactive inferenceAI consciousness

Read issue

Issue 18/August 22, 2026/10 entries

Adaptive harnesses, concept audits, and the measurement of agent research

Recent work tightens the loop between agents and their evaluation environments while exposing how easily claimed gains dissolve under better controls. The through-line runs from co-evolving validation tasks and programmable environment plugins to concept-level unlearning, evidence arbitration, and end-to-end formal theoretical research.

agentsbenchmarksautoformalizationco-evolution

Read issue

Issue 17/August 21, 2026/10 entries

Skill induction, adaptive allocation, and auditing self-improvement in agents

Recent work measures whether agents can truly improve their own training algorithms or skill memories, while formalizing routing, compute allocation, and transfer under tight budgets and noisy evaluation. The selection stresses controlled benchmarks, null models, and bicriteria guarantees over raw capability claims.

agentsbenchmarksmemoryadaptive reasoning

Read issue

Issue 16/August 20, 2026/10 entries

Credit assignment and strategy lock-in in long-horizon agents

Recent work converges on why multi-turn and agentic training stalls: outcome rewards starve the decisions that matter, post-training agents lock into an initial strategy, and fixed environments stop expanding the goal distribution. The selected papers introduce reverse-turn and dual-channel credit, group-calibrated distillation, learnable environment designers, and monitors for latent collusion, alongside theoretical treatments of preference indeterminacy and test-time exploitation failure.

credit assignmenton-policypost-trainingRL

Read issue

Issue 15/August 19, 2026/10 entries

Self-improving agents, reward hacking, and long-horizon trust

Fresh results expose variance and order sensitivity in memory-based self-improvement, show debate curbing judge hacking in RLAIF, and introduce monitors for ontological drift plus neurosymbolic world models. Complementary work tightens LLM judging under risk controls, surfaces MoE hallucination signals, and questions the self-consistency of preference estimates, while an essay reframes mathematical values under capable AI.

agentsevaluationLLMsself-improvement

Read issue

Issue 14/August 18, 2026/10 entries

Agent collectives, incremental memory, and grounded governance

Fresh work probes how language-model agents interact and coordinate at scale, schedules memory capacity for long contexts, and builds stricter mechanisms for compliance, planning specifications, and knowledge stores that agents themselves write.

multi-agentagenticagentsarchitecture

Read issue

Issue 13/August 17, 2026/10 entries

Recoverable state, world models, and long-horizon agency

Recent work tightens the loop between what agents remember, how they repair or rewrite their own models of the world, and how long-horizon research or evaluation can continue without collapse. Formal treatments of handover, explosion dynamics, and evidence aggregation sit alongside systems that checkpoint, twin, or re-anchor executable research state.

adaptive-stoppingagentsAI-R&DARC-AGI

Read issue

Issue 12/August 16, 2026/10 entries

Curriculum sandboxes, draft trees, and intent-group safety

Recent papers isolate knowledge exposure via grade-level pretraining, measure trajectory-level data influence without downstream tasks, verbalize sparse autoencoder features from decoder directions, probe Gricean specificity retreats at knowledge boundaries, construct diffusion-AR draft trees for speculative decoding, virtualize KV tokens for reclaimable caches, and supervise refusal by intent groups rather than surface form. Supporting results cover long-horizon Earth observation reasoning and open hierarchical models trained only on permissible data.

knowledge-boundariespretrainingattributionboosting

Read issue

Issue 11/August 15, 2026/10 entries

Reasoning definitions, moral divergence, and governed memory for agents

Recent work sharpens operational notions of reasoning and alignment while exposing gaps between label agreement and underlying moral grounds or integrity under pressure. Parallel systems papers advance associative recollection, source-bound persistent state, causal world models, and scientific agentic foundations for long-horizon reliability.

agentsalignmentevaluationposition

Read issue

Issue 10/August 14, 2026/10 entries

Long-horizon agents, repository-scale verification, and multi-agent failure modes

Fresh results push end-to-end AI scientists and interactive world models while new benchmarks and red-team experiments expose where agents still fail at coherent research, formal proof, command boundaries, and peer coordination. Early persona shaping and safety design trade-offs round out the picture of what must be built into systems before scale.

AI safetybenchmarkscoding agentsmulti-agent systems

Read issue

Issue 09/August 13, 2026/10 entries

Test-time harnesses, simulator collapse, and context trade-offs

Recent work examines how capability can be transferred or recovered at inference without parameter updates, how single-simulator training fails to generalize in multi-agent settings, and how longer contexts or compression choices reshape parametric knowledge, rankings, and memory reliability.

agentsbenchmarksreasoningRL

Read issue

Issue 08/August 12, 2026/10 entries

Catastrophic remembering, skill compression, and structured agent memory

Fresh results examine why agent prompts and skills grow without bound, how to compress reusable structure without costly evaluation, and how tree-structured or persistent memory supports self-correction, cross-run transfer, and long-horizon mathematical and search work.

interpretabilityagent memoryagentic codingAI-for-math

Read issue

Issue 07/August 11, 2026/10 entries

Internal probes, mutable harnesses, and judgment-free flow control

Recent work repeatedly shows that verbalized confidence, static retrieval, and pure content review are insufficient for reliable agents. The selected papers instead develop internal linear probes, external mutable document environments, representation-based intrinsic rewards, baked-in interpretability, coalition routing, and load-aware governance that bypasses semantic judgment altogether.

efficiencymulti-agentreasoningagent-design

Read issue

Issue 06/August 10, 2026/10 entries

Agent memory lifecycles and long-horizon competence

Recent papers examine how LLM agents handle evolving knowledge, refine procedural skills, and maintain reliability across extended interactions. They introduce mechanisms for conflict-aware retrieval, stale-memory revocation, proximal skill evolution, direct long-horizon prediction, and multi-turn clinical training.

agentsmemorymulti-turnreinforcement-learning

Read issue

Issue 05/August 9, 2026/10 entries

Harness optimization, selective routing, and self-distillation under sparse feedback

Recent work sharpens how systems around base models are improved: expensive harness search, when routing is learnable, Markov layer programs, recursive long-horizon task synthesis, and pivot-guided or non-semantic distillation. The shared thread is that gains now come less from raw scale and more from disciplined control of context, path, and supervision signals.

agentsefficiencyevaluationsynthetic data

Read issue

Issue 04/August 8, 2026/10 entries

Self-correction loops and the architecture of agent improvement

Recent work converges on how models and agents detect, localize, and repair their own failures: unsupervised on-policy distillation from internal consistency, weak probes that patch strong-model reasoning bugs, externalized state for long-horizon execution, and empirical signs that AI is already accelerating its own development.

agentsreasoningRLabstraction

Read issue

Issue 03/August 7, 2026/10 entries

World rehearsal, trajectory debugging, and selective trust in long-horizon agents

Fresh results on training and auditing LLM agents emphasize internalizing environment dynamics, tracing critical errors across long trajectories, adaptive on-policy self-distillation, and causal checks that reveal when tools or context truly drive decisions rather than create illusions of competence.

agentsevaluationtool-usereinforcement-learning

Read issue

Issue 02/August 5, 2026/10 entries

Embedded agency, continual skill evolution, and the mechanics of test-time reasoning

Fresh results probe how foundation-model agents cooperate via similarity inference, consolidate experience into transferable skills, and allocate inference compute, alongside formal treatments of causality, ablation, and pretraining inductive biases.

agentsbenchmarkcausalityinference

Read issue

Issue 01/August 4, 2026/10 entries

Agent memory continuity, failure detection, and rigorous evaluation of long-horizon reasoning

Recent work advances persistent state and hierarchical memory for agents and sequence models, develops efficient monitors and detectors for mid-run failures and cross-session misuse, and exposes how answer-only scoring inflates claims of scientific reasoning while formal synthesis and test-time latent optimization offer more reliable paths forward.

policy synthesisagent misuseagentic AIAI safety

Read issue