The headline: Output contracts got serious
Starting around v1.113, every agent in the system got standardized output schemas. Blocked status, retry reasons, revision findings, confidence scores — all structured, all predictable.
Why? Because when your orchestrator makes decisions based on agent output, ambiguous responses break things. A blocked_reason field beats a paragraph of prose every time.
The terminology shifted too: execution became workflow state, needs_revision became needs_retry. Cleaner names for cleaner mental models.
Fast path for trivial work
Not every task needs a full planning cycle. v1.113 introduced the fast path — a direct route for single-owner, bounded, low-risk tasks that skip the planner and reviewer entirely.
If you're fixing a typo or renaming a variable, you don't need a wave plan. The orchestrator now recognizes that and routes straight to the specialist. Promotion to a persistent plan happens automatically if scope creeps during execution.
Verification boundary
The orchestrator stopped second-guessing specialists. v1.117 added explicit rules: the orchestrator never re-verifies, re-tests, or re-analyzes completed work. That's the specialist's job.
This sounds obvious, but it prevents a real problem — double-verification waste and contradictory conclusions when two different agents evaluate the same code.
YAGNI in the planner
v1.114 enforced YAGNI principles inside the planner itself. Plans now actively resist scope creep during the planning phase, not just during implementation. Over-engineering gets flagged before it reaches the codebase.
Combined with the reviewer's critic mode (added in v1.105), there's now a two-layer defense against building things you don't need.
Critic mode for reviewers
Speaking of critic mode — it's not just for code review anymore. The reviewer agent can now evaluate decisions, not just implementations. Architecture choices, tradeoffs, technical direction — all get the same rigorous treatment.
The critic doesn't write code. It reads context, evaluates evidence, and produces a verdict. That separation of concerns matters.
Anti-slop, built in
v1.121 added anti-slop as a documented, built-in feature. No more AI-generated buzzword paragraphs or template filler in documentation. Every section exists because the product needs it.
The code simplifier agent also got rules to strip dead comments and reduce noise. The system now actively fights the kind of output that makes AI-generated code feel... AI-generated.
Quality directives everywhere
v1.122 pushed quality directives into every agent's rules. Standardized output formats, cleaned-up learn fields, consistent formatting across all agents.
This is the kind of work that doesn't show up in a feature list but makes the whole system more reliable. When every agent follows the same conventions, the orchestrator can trust their output without extra validation.
Performance
The docs now include real numbers: sub-$0.001 per API call on 100K+ token contexts, 82.8M+ tokens processed across 666 agent runs, typical responses in 2-5 seconds.
Prompt caching turns large contexts into sub-penny operations — a 10x cost reduction versus uncached input. That's not marketing; that's measured during Gem Team's own development.
The pattern
If you step back, the last 30 releases tell one story: maturity. The system went from "cool agent framework" to "engineering process that happens to use AI."
Every change reinforces the same principle: AI coding should be predictable, efficient, and self-correcting. Not because the AI is perfect, but because the process around it catches the mistakes.
That's the real feature.
Get started:
apm install mubaidr/gem-team --target copilot,claude,cursor,opencode,codex,gemini,windsurf