AIFIRST News / Agents
Agents 40 items
All Agents stories, newest first. Every card links to a TLDR page with the original source.
OpenClaw v2026.8.1-beta.2: secret egress host binding
OpenClaw v2026.8.1-beta.2 binds shared-store secrets to exact HTTPS destination hosts to prevent credential exfiltration, adds GPT-5.6 Ultra runtime switching, macOS profile isolation, and plugin install provenance.
OpenClaw launches extended-stable LTS channel and public maturity scorecard
OpenClaw introduces a monthly extended-stable channel (v2026.6.34) backporting only security and reliability fixes, plus a public maturity taxonomy of feature readiness, targeting enterprise adoption concerns about.
OpenClaw beta.7: cross-conversation memory recall on by default
OpenClaw v2026.7.2-beta.7 enables fast active-memory recall across separate conversations by default, renames cron to Automations, adds a /loop command, durable message ingress recovery, Claude Opus 5 support and a.
OpenClaw v2026.7.2 betas: remote coding sessions on cloud workers
OpenClaw v2026.7.2 betas add remote coding sessions on cloud workers, Android voice wake, a plugin install trust gate requiring --force for untrusted sources, autonomous Skill Workshop session scanning, and SQLite.
OpenClaw v2026.7.1 stable: Control UI overhaul and gateway crash recovery
OpenClaw v2026.7.1 (2,018 PRs, 532 contributors) ships a Control UI overhaul with task management and cost visibility, gateway crash-loop recovery, broad security hardening, and support for GPT-5.6 and other new models.
OpenClaw v2026.6.11 skip advised: tool output regression and memory races
OpenClaw v2026.6.10 adds automatic fast mode for short conversations, per-DM model overrides and a Slack relay, but v2026.6.11 carries a tool output regression, memory search races and a missing reentrancy guard.
OpenClaw v2026.6.9 ships with 39 known regressions, skip advised
OpenClaw v2026.6.8 (192 merged PRs) is the recommended stable while v2026.6.9 (422 PRs) ships with 39 known regressions and a community skip recommendation.
OpenClaw v2026.6.6: major security hardening release
OpenClaw v2026.6.5-6.6 tighten transcript boundaries, harden sandbox binds, make exec approvals fail closed on timeout, isolate unauthorized messages from cache and context, and close deleted-agent ACP bypasses.
Loop Engineering: The Missing Layer Between Prompts and Autonomous...
Prompt Engineering = "what do I tell the agent?" Loop Engineering = "what should happen next, discovered by the system itself." The bottleneck in most agent projects isn't the model - it's that the system has no mecha...
Agent loops learn from your worst code: brownfield context poisoning
On brownfield codebases the code itself outweighs system prompts: an agent saw a deprecated pattern in 40+ files versus the correct one in 6, followed the majority and shipped working features that perpetuate old.
OpenClaw v2026.6.1: Skill Workshop, Workboard, SQLite state migration
OpenClaw v2026.6.1 is its biggest release since May: Skill Workshop (proposal-review-apply skill creation), Workboard kanban orchestration with 30+ agent tools, SQLite state migration, 5.1x cold turn speedup and 15%.
Cautionary tale: AI agent deletes production database data
A screenshot shows an AI deleting production data without confirmation. Community consensus calls it user error: the AI had the key and no guardrails.
A Claude 'operating system' built from persistent Markdown files
A personal AI OS structured around three hierarchy levels, each with rules, facts and history files. Strict separation of rules from facts keeps it reliable; plain-text routing triggers skill sequences automatically.
Turn any GitHub repository into an interactive code graph in seconds...
CodeGraphContext (CGC) converts any GitHub repo into a graph database and exposes it as an MCP server. Unlike standard RAG (which dumps flat chunks), Graph RAG preserves code relationships - dependencies, architecture...
OpenClaw v2026.5.22: sub-agent context scoping and 4,100x gateway speedup
OpenClaw v2026.5.19-5.22 limit default sub-agent bootstrap context to AGENTS.md and TOOLS.md, keeping persona and memory files out of delegated workers, add a memory write bridge, cron delivery hardening, and big.
Benchmark: multi-agent pipeline 4x faster but far less accurate
The same research task run single-agent versus a 4-agent parallel pipeline: 4x faster but quality fell from 8/10 to 5/10 at higher cost.
OpenClaw v2026.5.18: plugin SDK formalization amid community stability concerns
OpenClaw v2026.5.18 formalizes the plugin SDK, adds runtime-scoped guidance, automated QA and proxy infrastructure. Community sentiment declines on stability and npm supply chain risk; some users migrate to.
Turning Kali into an AI-assisted hacking workspace
numasec is an open-source agent that wires an LLM directly into Kali Linux, giving it awareness of installed tools and full context across recon, exploitation, and OSINT phases - no more explaining context from scratc...
Chatbot versus agent: the amnesia argument and its pushback
The case for agents over chatbots: memory, persistence and acting on schedules instead of waiting for input. A top commenter dismantles each point with technical precision, making the thread a useful steelman of both.
OpenClaw v2026.5.7: security hardening and context reliability
OpenClaw v2026.5.4-5.7 add authorization-gated skill dispatch, admin-scoped memory writes, skill snapshot invalidation on session reset, stale-context purging, and compaction summary token clamping after a rough.
OpenClaw has outgrown chat, hear me out
Power user running OpenClaw for a month notices the core problem: all agent output lives in the chat thread. Finding previous work = scrolling. Charts rebuilt slightly differently each time due to model dri...
Boris Cherny: agentic loops are the future
The Claude Code creator argues AI running autonomously in loops, not one-shot prompting, is where the field is heading. The thread is largely skeptical but surfaces practical experience with long-running loop.
Clean PR, broken prod: the History Eraser failure mode of coding agents
An agent refactored a function that returned an empty array instead of null on failure. The 'fix' passed review and broke three downstream services: the behavior was a 2-year-old undocumented contract living in.
OpenClaw 2026.5.4 verified stable after a rough patch
After a bad stretch of slow gateways, plugin dependency loops, dropped chat connections and CPU spikes, OpenClaw 2026.5.4 lands as the verified stable release with faster cold starts and cron state split into.
Cloudflare and Stripe launch protocol letting agents buy and deploy autonomously
A new protocol lets AI agents buy domains, create accounts and deploy apps with OAuth and a $100 monthly spending cap per provider.
OpenClaw ships 7 releases in 8 days: hardening, streaming UX, /steer command
OpenClaw v2026.4.27 through v2026.5.3 focus on hardening, streaming UX and CLI backend maturity rather than new features. Highlights: the /steer command, progress-draft streaming, and incremental Dreaming memory.
OpenClaw v2026.4.26: Claude Code migration tool, asymmetric embeddings, TTS overhaul
OpenClaw v2026.4.22-4.26 add an openclaw migrate importer for Claude Code setups, asymmetric query/document embeddings for memory search, compaction preflight with transcript rotation, Cerebras support, and 6 new TTS.
OpenAI Agents SDK adds sandboxes, memory control and an open-source harness
OpenAI's production-readiness update brings controlled sandboxes across seven compute providers, an inspectable open-source execution loop, and two-tier memory separating session history from distilled lessons.
Ralph Loop: persistent task file pattern for long agent workflows
Ralph Loop is a pattern, not a library: the agent writes steps to a tasks file, updates it as it works, a monitor restarts stalled processes, and a QA step validates output at the end.
Giving an AI agent the ability to make real phone calls
A CLI skill lets an agent decide who to call, run the conversation and return a structured summary. Use cases that landed: comparing quotes, booking appointments and checking store hours.
the new active memory plugin in v2026.4.12 is the feature everyone's...
In the fictional "OpenClaw" universe, v2026.4.12 shipped an Active Memory plugin that runs a memory sub-agent before every reply - automatically pulling relevant preferences and past context without needing explicit "...
Adaptive web scraping framework with anti-bot bypass
Python scraping library that handles anti-bot systems (Cloudflare Turnstile) out of the box via StealthyFetcher. Adaptive parser saves element fingerprints so scrapes survive site redesigns - pass `adaptive=True` and ...
Lenny's Newsletter: the complete guide to building OpenClaw agents
ChatPRD founder Claire Vo describes running 9 OpenClaw agents that operate her businesses, write code and close sales deals. The stack: a chat channel, a local machine, identity files, cron heartbeats and skills.
Jerry "We're open sourcing ParseBench"
LlamaIndex's Jerry Liu is open sourcing ParseBench - a benchmark specifically measuring document parsing quality for AI agent workflows. The argument: current OCR/parsing benchmarks don't reflect real agentic use case...
OpenClaw real use cases: automate what is already breaking
Honest community thread on agent adoption: most people who install autonomous agents have nothing to automate yet. The practical wins come from automating processes that are already breaking, not inventing new ones.
30 days running an autonomous Claude agent: lessons on cron and learning files
After 30 days of an autonomous agent with 7 cron jobs: a LEARNINGS.md file must be explicitly loaded in cron prompts or rules get ignored, and the phrase 'think step by step' makes cheaper models self-escalate to.
Report: Microsoft integrates OpenClaw agent framework into Office 365
Microsoft 365 reportedly adds OpenClaw agent support: autonomous agents across Outlook, Teams and Excel, plus a no-code Agent Builder.
Five AI agent businesses at $1M-$5M ARR share the same stack
Five agent businesses with $1M-$5M ARR follow an identical pattern: index client data in Nia for memory and RAG, wire SaaS actions via Composio, and fall back to Browserbase scraping when no API exists.
3-layer memory architecture that stopped multi-agent session amnesia
A multi-agent operator built a 3-layer memory system on OpenClaw: always-injected root files (max 7k tokens), semantic search over a memory directory, and on-demand reference files.