Three weeks since the June roundup, and the changelog did not slow down: nine platform releases, v0.16.2 through v0.17.7. But the bigger story this time is not a feature list. Everruns stopped being one repository. There are now two more open-source projects in the family — Yolop, a terminal coding agent, and Mira, an evaluation framework — and much of the platform work points at one destination: opening Everruns Cloud to the public.
Here is the map, then the details. Every platform feature links into the documentation or its release notes.
Two new projects
Yolop
Yolop is a new terminal coding agent: one binary that plans, edits, runs, and verifies code in your repository, with persistent sessions, agent skills, MCP servers, and editor integration over the Agent Client Protocol. It is built on the everruns-runtime we wrote about embedding in May, so its runs are durable and survive a restart out of the box.
It stays out of your way. Yolop runs autonomously by default — writes, edits, and commands without a pop-up per action — behind a standing write blocklist, and asks only when a step actually warrants it: soft approvals are AI-judged consent points with an audit trail, not a dialog for every file write. (The name is Ukrainian — йолоп: a dummy, a fool — describing the attitude, not the results.)
Six versions in, it handles multi-language repo maps, native git worktrees, background execution, opt-in LSP-backed code navigation, structural edits via an ast-grep-backed ast_edit tool, and authenticated MCP server management.
See the Yolop page or the source. Installing it is one line:
brew install everruns/tap/yolop # or: cargo install yolop --lockedMira
An agent that cannot be measured regresses silently — a model swap, a prompt tweak, a runtime upgrade, and nobody notices until a user does. Mira (Ukrainian міра: measure, standard) is how you catch it: a Rust-first, code-first evaluation framework for multi-turn, tool-using, long-running agent trajectories. You write evals in code — Rust natively, or Python and TypeScript over a documented protocol — and one mira host CLI runs them across a model matrix, scores the results, and reports, with saved and resumable runs and CI-native output including a self-contained HTML report.
You score more than the final answer. Alongside deterministic checks (contains, regex, tool_called), Mira has trajectory-structure scorers that judge how the agent worked, operational budgets (tokens, cost, latency), combinators, and LLM-as-judge — composed freely into one verdict, so a regression in cost or tool use fails the run as loudly as a wrong answer does.
Mira is at v0.4.0; the crate is mira-eval.
Everruns Cloud is coming
Everruns has been self-hostable open source from day one. The next step is the one people keep asking for: a managed Everruns you don’t run yourself. We are preparing to open Everruns Cloud — the same platform we already operate at app.everruns.com — to the public, so you can point an agent at it and go without standing up Postgres, a worker, and a queue.
No date to announce yet — join the launch list and we will email you the moment it opens, with occasional progress along the way.
Sessions that do more
Real-time session tasks
Session tasks got the June treatment of webhooks and kept going. Task state transitions now wake a running session mid-turn — on the durable path, so a wake survives a crash like everything else — every transition is delivered rather than just the terminal ones, webhook configs are per-task, and structured task results are exposed over both A2A and MCP (v0.17.7). An agent waiting on background work no longer polls for it; the work interrupts the agent.
Forked, detached, and handoff sessions
Sessions learned three new tricks. Forking copies a session into an independent branch from any point in the conversation — cheap parallel experimentation on a real history (v0.17.1). Detached peer sessions spawn alongside the parent without blocking it, and handoff agents can be invited into an existing session to take over (v0.17.7).
Background subagents
spawn_subagent now runs in the background by default, waking the parent on completion (v0.17.5). Combined with June’s durable subagents, a parent can fan out work, keep going, and pick up results as they land — the same shape as the session-task wakes above, applied to child agents.
Scoped memory
Memory stores were org-scoped; they can now be scoped to a single agent or a single user (v0.17.7) — an agent’s accumulated knowledge and a user’s preferences no longer have to share one org-wide pool.
Evals and trajectories
Observers
June’s online judges grew a face. Observers — the machinery that scores live sessions — now have their own UI: create and edit them against a scorer catalog, review results in a per-observer Quality tab, and llm_judge scorers that reference a model your org cannot use are rejected up front instead of failing at runtime (v0.17.2).
Eval runs: compare, share, import
Three quality-of-life pieces for evals: side-by-side run comparison with regression highlighting (v0.17.4), read-only revocable share links for runs (v0.17.5), and ingestion of externally-executed runs with full attribution, transcript, and matrix views (v0.17.2). That last one is the Mira bridge: run an eval anywhere, view it in the platform.
ATIF across the stack
One thread ties all three projects together this cycle: the Agent Trajectory Interchange Format. The platform exports sessions and eval datasets to ATIF, segmented for large sessions (v0.17.6); Yolop writes ATIF trajectory files with --trajectory-out; Mira exports saved runs to ATIF with mira export --format atif, verdicts stamped in. One format for what an agent actually did — ready for fine-tuning, RL, or trajectory-visualization tooling, whichever project produced it.
Also since June
Smaller changes worth knowing about:
- MCP, always. The MCP endpoint is now always exposed — the feature flag is gone (v0.16.2) — and it conforms to the 2026-07-28 stateless MCP spec (v0.17.1).
- Parallel tool calls. A
parallel_tool_callspreference, unified across all LLM drivers (v0.16.2). - Providers and models. Typed multi-field credential schemas make OAuth a first-class way to connect a provider, starting with OpenRouter from the Add Provider dialog. New model profiles: Claude Sonnet 5 and GPT-5.6 Sol, Terra, and Luna.
- Streaming reconnect. LLM streaming responses reconnect automatically on transport failure (v0.17.6) — one less way for an hours-long run to die.
- UI polish. A unified five-zone layout across the main pages, a combined model-and-effort composer menu with recents, an LLM speed selector mapped to OpenAI’s
service_tier, and keyboard turn navigation for chat transcripts.
All of it is open source, MIT-licensed: everruns/everruns, everruns/yolop, everruns/mira. The reference is at docs.everruns.com. And if a managed Everruns is the thing you have been waiting for — contact@everruns.com, we would like to hear what you plan to run on it.