Engineering notes on agent orchestration, cost control, security, and shipping with Swarm 357.
Status26 July 2026 by Alex Cinovoj
The July 2026 maturity picture in plain language. What hardened this year, what is still Beta and why, and the four things that are explicitly not implemented.
Scenarios20 July 2026
Triage, macro drafting, and escalation summaries with no shell access at all. What the support layer does without it, and when a human unlocks a command.
Architecture13 July 2026
How six business layers and 357 roles were drawn, what a soul template holds, and the three mistakes that make a role catalog useless.
API6 July 2026
Streaming agent events over server-sent events, authenticating writes without breaking the stream, and publishing telemetry without leaking task content.
Scenarios29 June 2026
A worked scenario for research, drafting, and handoff notes, with the send left to a human. What the sales layer refuses to do, and why that line matters.
Operations22 June 2026
A run that dies at step nine should not restart at step one. How SQLite-backed state turns a crash into a resume, and why cancel must be a persisted flag.
Cost control15 June 2026
Point the runtime at OpenRouter, use a cheap tool-capable model to iterate, and keep premium models for baselines. Why free models break tool calling.
Cost control8 June 2026
Reserve before the call, budget by layer not by run, degrade loudly, and cap the context you load. Patterns that keep parallel agents inside your number.
Comparison1 June 2026
A state graph gives you explicit control flow. A role catalog gives you governed workers. When you need one, when you need both, and how to tell.
Definitions26 May 2026
The word swarm gets used for three different systems. A definition that survives a security review, plus the four vendor claims that do not hold up.
Deployment12 May 2026
Split the API from the landing site, keep write keys server-only, force simulate on the public demo, and let the container make Bash fail closed.
Evals28 April 2026
Single-agent and swarm metrics measure different systems. An eval methodology with budget caps, baselines in version control, and the numbers that lie.
Comparison14 April 2026
Ad-hoc crews, conversation graphs, and pre-declared role catalogs solve different problems. A criteria table, and when a catalog is the wrong answer.
Security24 March 2026
A durable ApprovalRecord, a control plane on both CLI and HTTP, and timeouts that reject. Putting a person in front of side effects without stalling runs.
Security10 March 2026
Thirteen patterns, argv-level validation, and a deny-by-default posture in production. What a bash policy gate stops, and the three things it will never stop.
Memory24 February 2026
Flat-file topics for memory you read, a single-file .mv2 archive for memory you search. How the Memvid bridge gives agents recall you can copy and diff.
Cost control10 February 2026
Layer budgets, an atomic ledger, and a logged downgrade at 80 percent. The FinOps primitives that keep a multi-agent run from becoming a surprise invoice.
Getting started27 January 2026
Install the package, run the simulated demo with no API key, then make one live call on a cheap model. With the four errors people actually hit.
Orchestration13 January 2026
357 roles is an org chart for agents, not 357 models at once. How role catalogs, layer routing, and hard agent caps keep a multi-agent system solvent.
Subscribe via RSS.