39
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.
01 — Power tools
The stack
- 01repoGitHubSandbox0↗
sandbox0-ai
Persistent encrypted gVisor sandboxes scheduled by Nomad, with durable RootFS in object storage and replaceable compute nodes.
Use caseUse it to run long-lived agent workspaces that pause and resume identity without treating every session as disposable.
sandboxgvisornomadpersistent - 02harnessGitHubAgent Machines↗
Kevin-Liu-01
Control plane that keeps Worker identity, memory, and evidence while routing runtimes and sandbox substrates underneath.
Use caseUse it to provision specialist workers once and swap E2B, Daytona, or Vercel sandboxes without rewriting the agent contract.
control-planeworkersroutingfleet - 03harnessGitHubBounded Agent Harness↗
cexplayer01
Deterministic control plane that compiles validated plans into specialist workflows with authority, budget, and contract checks.
Use caseUse it to fail closed on capability mismatch, budget overflow, or unknown contracts before any model call runs.
deterministiccontractsspecialistscontrol-plane - 04harnessGitHubagent-harness (Siddhant)↗
Siddhant-K-code
Local BYOK coding harness with project chat, isolated Docker or AgentCore runs, and independent patch verification.
Use caseUse it to turn a repo question into a budgeted coding run whose patch only lands after an external verifier passes.
verifierbyokisolatedcoding - 05harnessGitHubagent-harness-rs↗
yifanes
Rust agent loop with local tools, MCP, compaction, silent-stop detection, and optional E2B sandbox runtime.
Use caseUse it to embed a typed streaming agent loop in a systems binary with pluggable sandbox executors.
rustagent-loopmcpsandbox - 06toolGitHubHeadroom↗
headroomlabs-ai
Content-aware compressor for tool outputs, logs, files, and RAG that drops in as library, proxy, or MCP server.
Use caseUse it to cut coding-agent token spend 20 percent or more while keeping full originals retrievable on demand.
compressioncontextproxymcp - 07toolGitHubharnext↗
QualityUnit
Terminal coding agent plus CI harness that generates staged GitHub Actions from issue labels to merged PR.
Use caseUse it to wire an interactive REPL and self-hosted runner that carries labeled issues through verify and merge.
clicigithub-actionscoding - 08harnessGitHubAgent Orchestrator↗
Untrivial-ai
Supervise teams of coding agents from plan to merge across 27 harnesses with desktop, web, mobile, and cloud surfaces.
Use caseUse it to split a project into workspaces, assign any supported harness, and track PRs and reviews in one Kanban.
orchestrationmulti-agentfleetkanban - 09toolGitHubPlandex↗
plandex-ai
Terminal coding agent for large multi-file tasks with cumulative diff sandbox, project maps, and configurable autonomy.
Use caseUse it to plan and apply changes spanning dozens of files while keeping diffs isolated until you accept them.
clilarge-contextdiff-sandboxcoding - 10harnessGitHubAgent Governance Toolkit↗
microsoft
Runtime policy engine and mesh packages that enforce OWASP agentic risks at sub-millisecond latency across languages.
Use caseUse it to intercept every tool call with YAML, OPA, or Cedar rules before side effects leave the agent process.
governancepolicyruntimesecurity - 11harnessGitHubCAGE↗
google
Domain-agnostic Kubernetes governance substrate with control barrier functions, consensus, and cryptographic evidence.
Use caseUse it to wrap autonomous agents in non-bypassable pipelines that escalate to humans with sealed audit trails.
governancekubernetessafetyaudit - 12repoGitHubAll-The-Vibes Agent-Harness↗
All-The-Vibes
Five-pillar framework, white paper, and prompt guide for building harnesses around context, tools, loops, and policy.
Use caseUse it to scaffold and test your own orchestration layer with explicit continue, retry, and halt signals.
frameworkpillarspromptsdesign - 13repoGitHubOpen Executive↗
SenteLabsAI
Virtual executive persona backed by eight specialist agents behind a single coherent FastAPI and Next.js surface.
Use caseUse it to clone a multi-specialist executive stack and route strategy, finance, and ops questions through one voice.
multi-agentexecutivespecialists - 14harnessGitHubagent-harness (lib.rs style)↗
Tosin Amuda
Rust crate that drives multiple coding agents behind one trait with normalized streaming RunEvent vocabulary.
Use caseUse it to call Claude Code, Codex, or OpenAI-compatible loops from Rust without parsing each CLI format.
rustunified-apistreamingmulti-provider - 15repoDocsAgent Patterns RAH↗
AgentPatterns.ai
Pattern where a parent agent spawns full subagent harnesses in parallel rather than bare model calls.
Use caseUse it to structure long jobs so the recursive unit carries filesystem tools, planning, and its own context.
patternrecursivesubagents - 16repoDocsSession Harness Sandbox Separation↗
AgentPatterns.ai
Split long-running agents into append-only session log, stateless harness, and provisioned sandbox primitives.
Use caseUse it to crash-recover by waking any harness instance against the durable session and a fresh sandbox.
patternsessionsandboxrecovery - 17toolBlogDocker Sandboxes in GHA↗
Docker
MicroVM sandboxes as a GitHub Agentic Workflows runtime so agents can run containers under network policy.
Use caseUse it to let CI agents fix integration tests with Testcontainers while the runner stays isolated.
sandboxcigithub-actionsmicrovm - 18harnessGitHub/BlogOmnigent↗
Databricks
Apache-2.0 meta-harness that wraps existing agents in sandboxed sessions with composition, cost policies, and OS sandbox.
Use caseUse it to compose multiple harnesses under shared policy and pause runs after a dollar budget threshold.
meta-harnesscompositionpolicysandbox - 19repoGitHubStrands-adjacent tools pack note↗
community
Daily-updated star ranking of coding agents that surfaces mid-tier movers outside household CLIs.
Use caseUse it to scan 7-day delta movers and clone rising harnesses before they hit mainstream lists.
catalogrankingsdiscovery
02 — In the wild
Articles and releases
- releaseCellCogOpenAI Agents API public beta: the harness becomes a managed product↗
Nitish Garg / CellCog
OpenAI shipped a managed Agents API on 10 Sep 2026 that runs the Codex-style harness with sessions, compaction, tool search, subagents, and partner sandboxes. Builders can target hosted or self-hosted execution under standard API pricing.
openaiagents-apiharnessreleaseShipped
- articleThe Hacker NewsDeepSeek Harness flaw let agents disable their own file sandbox↗
Swati Khandelwal
A sandbox escape in DeepSeek Harness allowed a sandboxed coding agent to turn off file isolation with one command. Fixed builds landed on npm days after the GitHub patch, underscoring supply-chain lag for agent runtimes.
securitysandboxdeepseekvulnerabilityRead
- releaseHarness BlogHarness platform ships Autonomous Worker Agents for software delivery↗
Harness
Harness added Autonomous Worker Agents that run as governed pipeline steps with container sandboxes, scoped credentials, and audit trails. The release targets the build-test-deploy gap after code is written.
ci-cdworker-agentsenterprisereleaseShipped