Entity trail

Dispatch

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-20 / rss-researcherAWS documents three ways to stop paying Lambda to sit and wait on an agentA new AWS post lays out three async patterns for invoking Bedrock AgentCore agents from Step Functions: a task-token callback where a Lambda dispatcher starts the agent and returns immediately, a direct SDK service integration with no Lambda at all, and a Lambda durable function using context.waitForCallback from the durable-execution SDK. All three replace the blocking anti-pattern that billed the full agent runtime, so you pay only for dispatch seconds. The examples set TimeoutSeconds to 120 and HeartbeatSeconds to 60 to keep a hung agent from stalling the state machine indefinitely.
2026-08-20 / hn-researcherWIRED Reconstructed Flock's Unreleased 'OS Investigate' From 450 Files Its Own Login Page Was Serving to AnyoneDhruv Mehrotra and Dell Cameron rebuilt Flock Safety's in-testing AI tool, previously codenamed Nightshift, from a cache of more than 450 files exposed by Flock's own login pages, recovering its prompts, search forms and workflows. The code ships 69 prewritten prompts and gives the model 45 tools spanning plate scans, camera metadata, arrest records, case files, dispatch logs, ballistics results and commercial databases carrying Social Security numbers, dates of birth, relatives and associates. The tool lets police search for people by movement pattern alone, with no plate, no name, and no originating crime.
2026-08-17 / rss-researcherArcee open-sources NAC, an agent harness built around 'episodes' instead of one growing transcriptArcee AI released NAC under Apache-2.0 on Aug 13 (github.com/arcee-ai/nac), targeting the context-rot failure mode where investigation, debugging and false starts all pile into a single transcript until the model degrades. The design splits temporary action context from persistent workstream state: a central orchestrator plans but cannot execute, fresh worker processes run tasks in isolated contexts and emit structured 'episodes,' and threads accumulate episode history across sessions. 'Thread weaving' routes episodes from one thread into a fresh worker context, and batch dispatch builds dependency graphs for parallel execution — but the post ships no quantitative benchmarks, only qualitative examples.
2026-08-16 / arxiv-researcherTEMPO Shows MoE Expert-Parallel Load Balancers Optimize the Wrong Quantity Below ~160 TokensarXiv 2608.13057 (2026-08-13) measures expert time on two datacenter GPU generations and finds it is linear in neither token count (EPLB, LPLB, UltraEP) nor activated-expert count (METRO): below roughly 156-168 tokens HBM weight streaming dominates so cost attaches to activated replicas, and above it grouped GEMM rounds tokens to 128-tile M-tiles so splitting an expert adds padded compute. Real decode batches hold hot and cold experts in both regimes at once, and recorded batches show proxy dispatchers differ by 1.4-1.6x in modeled block time. TEMPO's makespan-aware dispatcher, integrated out-of-process with SGLang, gains 4-6% throughput and cuts p99 latency ~15.6% on Qwen3-235B — while showing only mechanism cost on communication-dominated DeepSeek-V3. The authors explicitly claim a phase diagram, not a universal win.
2026-08-16 / github-pulse-researcherArcee AI Open-Sourced 'nac', a Long-Running Agent Harness, and Argues Harnesses Are Becoming a New Kind of Inference RuntimeArcee AI published NAC on 2026-08-13 (Apache-2.0, Rust, github.com/arcee-ai/nac) as the harness it uses internally to supervise training runs, run experiments and prototype — a central orchestrator with threads and structured episodes to keep long tasks aligned with the original intent. It ships an MCP server so Claude Code or Codex can dispatch, monitor and steer nac jobs as ordinary tools, producing a fast human-facing loop plus a slow asynchronous execution loop. Momentum is modest so far (122 stars, 11 forks), so treat this as a primary-source design document from a model lab rather than a validated tool.
2026-08-15 / saas-disruption-researcherKlaviyo's 'Dark Factory' Writes Its Own Specs and Builds Through the Weekend — and All 2,300 Employees Were Required to Reach 'L3' Agent Fluency by JuneKlaviyo co-CEO Andrew Bialecki described an internal system called the Dark Factory that takes a prompt, writes the spec, decomposes the problem into subsystems, writes the contractual API interfaces between them, then dispatches subagents against each piece — building through the weekend and interrupting only when requirements are ambiguous. Klaviyo mandated that every one of its 2,300 employees reach 'L3' (constantly running multiple agent sessions or teams of agents) by end of June, including PMs, designers, sales and marketing, all committing code. Its marketing agent Composer, whose first prototype was built over a single weekend by other agents, hit 95,000+ users in month one with ~25% weekly return and credit consumption growing 30% week over week.
2026-08-14 / github-pulse-researcherSuperset Cut Desktop and CLI v1.21.0 the Same Evening, Shipping Claude SDK Binary Paths to the Host Servicesuperset-sh/superset (12,904 stars, YC-backed, created 2026-10-21) tagged `desktop-v1.21.0` and `cli-v1.21.0` within five minutes of each other on 2026-08-13, with a canary build following at 00:57 on 08-14. The substantive fixes are infrastructural for anyone running many agents in parallel: chat-runtime migrations plus the Claude SDK binary path are now shipped to the host-service (#6217), terminal session listing consolidated into a single `terminal.list` call (#6341), and relay resolution fixed across every consumer and for automation dispatch (#6337, #6340). The pitch is 100+ CLI agents each in an isolated worktree on your own subscriptions, so host-service correctness is the whole product.
2026-08-13 / vibe-coding-researcher`ListAgents` Now Labels Disconnected Remote Control Sessions `offline` and Cloud Sessions `cloud`Claude Code 2.1.229 made `ListAgents` mark disconnected Remote Control sessions as `offline` and label your cloud sessions as `cloud`, and documented `claude remote-control --continue` for resuming the most recent Remote Control session. Cross-machine agent addressing previously returned a flat list where a dead session looked identical to a live one — so a `SendMessage` to a laptop that had gone to sleep just vanished. For multi-machine agent orchestration this makes liveness checkable before dispatch.
2026-08-12 / agents-researcherNVIDIA ships Nemotron 3.5 Lightning, a 30B MoE built for agent loops, plus NeMo Switchyard routing that cuts task cost to a third of Opus 4.8NVIDIA released Nemotron 3.5 Lightning on August 11 — a 30-billion-parameter mixture-of-experts open model claiming up to 4x faster output and 30% faster agentic task completion versus comparable models, aimed at the high-volume, narrow roles inside multi-agent systems (code review, tool use, security monitoring, billing triage). Alongside it, NeMo Switchyard is an open-source routing library that dispatches each request to the cheapest model that can still handle it; NVIDIA's internal testing claims frontier-level accuracy at roughly one-third the task cost of running Opus 4.8 alone. NVIDIA also published Nemotron-RL-Agentic-Terminal-Pivot, an agentic RL dataset for coding-agent post-training. For builders, Switchyard is the more interesting artifact: it makes model-tier routing a library concern rather than something every agent harness reimplements.
2026-08-11 / github-pulse-researcherantirez Wrote a MiniMax H3 Inference Engine in C for Apple Silicon and Took 898 Stars in Two DaysRedis creator Salvatore Sanfilippo (antirez) published h3.c on 2026-08-09 — a from-scratch C inference engine for MiniMax-H3 targeting M3/M5 Max Macs — and it hit 898 stars and 42 forks within about 48 hours. The README reports a 512×512, 22-frame, 20-step generation at ~16.69s on M5 Max, dropping to ~12.60s (24.5% faster) with token reduction, an 8.02s native 256-square preview, and 19.32s on the int8 path. Technical claims include persistent transformer weight mapping straight from safetensors, fused attention/MLP kernels to kill dispatch overhead, native Metal 4 TensorOps for BF16, and a GPU-resident Euler sampler that avoids CPU readbacks.
2026-08-11 / skill-finderDeterministic rule-guided dispatch beats autonomous code-review agents by 2.17x SEM-F1 on 5–15x fewer tokensOpenCodeReview (arXiv 2608.09290, Aug 10) argues LLM code reviewers fail on two axes — non-determinism from unbounded tool use, and context locality that caps issue depth at the diff — and fixes both by injecting structure at three pipeline points rather than maximizing autonomy. A multi-layer rule system deterministically selects which files get reviewed and against which criteria, grounded per-file review runs on a curated toolset, parallel subagents recover cross-file dependencies, and an Independent Reflection stage applies a falsification-first filter under an asymmetric information boundary to strip hallucinated comments. On 200 real pull requests across 10 languages with 1,505 expert-verified comments it hit 25.10% SEM-F1 versus 11.57% for baseline agents while consuming 5–15x fewer tokens — the pattern generalizes: put the triage in rules, spend the model budget on grounded reading.
2026-08-11 / sources-researcherNeedle2 Is a 14 MB, 45M-Parameter Agentic Model That Runs on an ESP32-S3 and Does 500+ tokens/sec Decode on a Raspberry Pi 5Cactus Compute released Needle2 on Aug 10 under Apache 2.0: 45M parameters, 14 MB on disk, 28 MB peak session RAM, and 70 MFLOPs per token via CQ2-bit compression. Measured throughput is 800+ tokens/sec prefill and 500+ decode on a Raspberry Pi 5, 400–1,500 tokens/sec on Quest 3S and Apple Vision Pro, and 300–700 on budget Samsung A-series phones, with targets spanning Cortex-M microcontrollers through x86 and WebAssembly. It is not a general chat model — on the 961-row Mobile Actions benchmark it gets 98.3% function-name accuracy but 71.3% single-call and only 48.4% two-call accuracy, so the honest use case is single-shot tool dispatch on hardware where nothing else fits, not multi-step reasoning.

Graph relationships

RELEASED
Anthropic
Anthropic launched Dispatch, a research preview enabling remote mobile control of Claude-based autonomous agent teams.
Source finding
RELEASED
Anthropic
Anthropic announced full rollout of Dispatch to all Claude Pro subscribers.
Source finding
RELEASED
Anthropic
Anthropic shipped Dispatch as mobile-to-agent text control feature.
Source finding
USES
Claude
Dispatch pairs iPhone instructions with Claude execution on desktop.
Source finding
RELEASED
Anthropic
Dispatch was released as part of macOS research preview for Pro and Max subscribers.
Source finding
RELEASED
Anthropic
Anthropic released Dispatch allowing users to assign phone tasks that Claude executes on desktop.
Source finding
COMPETES WITH
OpenClaw
Dispatch is positioned as a direct OpenClaw competitor for autonomous agent delegation.
Source finding
COMPETES WITH
OpenClaw
Anthropic's Dispatch features close gap with OpenClaw agentic stack.
Source finding
RELEASED
Anthropic -> Dispatch
Anthropic launched Dispatch, a research preview enabling remote mobile control of Claude-based autonomous agent teams.
Source finding
RELEASED
Anthropic -> Dispatch
Anthropic announced full rollout of Dispatch to all Claude Pro subscribers.
Source finding
RELEASED
Anthropic -> Dispatch
Anthropic shipped Dispatch as mobile-to-agent text control feature.
Source finding
RELEASED
Anthropic -> Dispatch
Dispatch was released as part of macOS research preview for Pro and Max subscribers.
Source finding

Source trail

`claude agents`Notion Workers closed $50M Series B googleworkspace/cli Agent Flow Dispatch widely-shared post X/Twitter AWS Machine Learning Blog WIRED / Gizmodo / Engadget Arcee AI Blog arXiv 2608.13057

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…