Swarm 357 mid-year status: what is Stable, Beta, and not built

Halfway through 2026, here is where Swarm 357 actually stands. Four maturity levels, no adjectives doing work that a table should do.

I publish this because I have been on the buying side of an agent framework whose README implied everything worked. Finding the gaps during an integration is expensive. Finding them on a status page is free.

The levels

Stable. Tested, documented, safe for production use under the stated constraints. Beta. Working but evolving. The API may change. Pin a version. Experimental. Functional but incomplete or opt-in. Do not build on it. Not implemented. Explicitly out of scope for this release. Not coming soon by default.

What is Stable

The pattern across that list: the pieces that hardened first are the ones where getting it wrong is unrecoverable. Security controls and the refusal to fake success came before orchestration polish. That was deliberate.

What is Beta

Swarm orchestration and execute_layer(), the atomic BudgetLedger, CostController, the Memvid .mv2 bridge, the HTTP API, HITL approvals, the SSE event stream, structured traces, UltraPlan, and the CLI verbs boot, run, inspect, resume, cancel, approve, reject, replay, fork, eval, and serve.

That is most of the interesting surface, and Beta here means the API may still move, not that it is untested. Concretely:

• Orchestration does structured JSON routing with durable RunState and SQLite checkpoints. It works. The routing heuristics are still changing.
• The budget ledger reserves and commits under an asyncio lock. The locking is sound; estimation accuracy varies by model and tool use, which is the honest reason for the label.
• The Memvid bridge requires the memvid-swarm-bridge binary and wraps crates.io memvid-core. The integration surface is still moving.
• HITL approvals carry a real Bash gate, durable records, and timeouts that reject.

Pin a version. Read the changelog before upgrading.

What is Experimental

run_dream_cycle(). It is heuristic memory consolidation, and consolidation quality is not guaranteed. I use it. I do not rely on it to prevent memory accumulation, and neither should you. A scheduled job that archives old topics is more trustworthy than a clever pass. See portable agent memory with Memvid.

What is not implemented

Four things, stated plainly because vague roadmaps waste everybody's time.

Multi-tenant SaaS control plane and SSO. There is no tenancy model. A shared API key is not one.

Distributed rate limiting across replicas. The limiter ships single-process. Stacking per-process caps across replicas is not the same guarantee, and if you need a hard global quota, this is not it yet.

Guaranteed dream-cycle consolidation quality. Covered above.

Full Opik cloud observability. Local JSONL is the source of truth. Structured traces land in .swarm/traces.jsonl, with optional OpenTelemetry export behind SWARM_OTEL_EXPORT=1. Cloud observability is not wired.

What hardened this year

Three areas moved most, and all three came from something going wrong first.

Cost. The reserve-and-commit ledger replaced post-hoc accounting after parallel agents in one layer raced past a budget. The 80 percent downgrade to Haiku now emits a model_downgrade event, because a silent quality change is a debugging tax. Detail in four token budgeting patterns for multi-agent systems.

Security posture in the container. Bash deny and HITL key off SWARM_SERVER_MODE, which the API image sets and the server re-asserts at import. The posture comes from the artifact rather than from a deploy checklist. The public demo forces simulation server-side, clamps the budget, caps the body, bounds the input, and rate limits per IP. Unauthenticated run reads are redacted rather than blocked.

Honesty about keyless runs. An early build returned plausible output without a key. That is the worst failure in this category, because everything looks fine and nothing is real. Keyless runs now refuse unless you explicitly ask for simulation.

How to read any agent framework's status page

Three questions I ask, including of this one.

1. Is there a level below Beta? A project with only Stable and Coming Soon is not measuring itself.
2. Does "not implemented" have entries? An empty list means nobody wrote down the boundary.
3. Do the published numbers have a committed artifact behind them? Eval baselines here live in evals/baselines/latest.json and every published figure is rendered from that file. If a number cannot be traced to a run, it is marketing.

Check it yourself

pip install techtide-swarm
swarm init
swarm demo
swarm eval

The maintained table is status and maturity. The verification scorecard lists what has been checked and how, evals methodology covers scoring, and the roadmap covers what is next. Source on GitHub, package on PyPI.