v1.7.0 — Trust & Safety section + protocol hardening

What's new in v1.7.0

A backward-compatible refinement of the AGENTS.md protocol, informed by an external best-practice review (the agents.md spec, OpenAI/Anthropic/OWASP guidance, and self-evolving-agent research). The meaning of existing rules is unchanged.

Trust & Safety — new top-level section

The trust/security model moves out of the Startup checklist into its own prominent section and gains:

• Explicit instruction precedence: the user's current message > this file > the nearest nested AGENTS.md; no instruction file overrides the safety, confirmation, and permission rules.
• Honest framing of injection defense: the priority tiers are a best-effort heuristic, not a security boundary — in-context confirmation plus the runtime's own permission/sandbox controls are the real enforcement layer.
• Executable guards: for any rule that must not be violated, prefer an executable guard (test, hook, sandbox, permission boundary) over text alone.

Hard Constraints

• Least privilege: use only the minimum credentials, tokens, tool access, and scopes the task needs.

Self-Evolution

• The numeric promote/demote thresholds are now framed as tunable defaults, not validated constants, with a human-gated fallback when no reliable outcome ledger exists.
• Added a per-session overfitting check so a single session does not overfit the protocol to one-off events.

Docs

• README (EN/ZH) security FAQ repointed to the new Trust & Safety section.

Both language templates (AGENTS.md, AGENTS.zh-CN.md) stay in sync (10 sections each).

---

Pin this release: curl -fsSL https://raw.githubusercontent.com/yeasy/agentgo/v1.7.0/AGENTS.md -o AGENTS.md

v1.6.0

Changes

• Add bounded self-evolution discipline for AgentGo rules, workflows, and skills.
• Record rejected candidate updates as negative feedback when they prevent repeated mistakes.
• Add transfer caution for reusing workflows and skills across models, harnesses, repository types, or task families.
• Sync English and Simplified Chinese docs and bump the template marker to v1.6.0.

Validation

• git diff --check
• Bilingual keyword and version scan

v1.5.0

AgentGo v1.5.0 adds lightweight protocol guidance for project memory and quality signals.

• Add risk-scoped testability and observability assessment for code, operations, and behavior changes.
• Add optional evidence-backed project relationship maps under .agents/memory/project-map.md for complex or repeated navigation work.
• Clarify in the README files that AgentGo does not require full knowledge graphs, automatic instrumentation, or command-by-command history logs.

v1.4.0 — Operationalized self-evolution lifecycle

Self-evolution moves from a lifecycle in name to a lifecycle in numbers.

Promote / demote / archive — now data-driven

• Outcomes ledger now requires agent field and result enum (helped | hurt | no_effect | corrected) so promotion and demotion thresholds can be counted mechanically instead of estimated.
• Promote a candidate only after ≥3 distinct uses with result=helped and no unresolved corrected/hurt in the last 5 uses.
• Demote an active capability when ≥2 of its last 5 uses are corrected/hurt, when it is unreferenced for 90 days, or when a health check flags it stale, noisy, or superseded.
• Archive only after maintenance confirms no active outcome still depends on it.

Promotion gate

• experiments/ → rules/ / workflows/ / skills/ promotion now requires user confirmation.
• Agent-authored entries in rules/, workflows/, skills/, experiments/ are documented as advisory standing context, not authoritative instructions; they must be cross-checked against current project artifacts before being followed.

Outcomes aging + rollback path

• Outcomes are archived alongside their referenced capability; entries older than 90 days that no longer point to an active capability become cleanup candidates so the ledger never outgrows what it serves.
• Harmful demotion now triggers re-review of dependent active outcomes and records affected artifacts in memory/open-items.md instead of leaving them silently in place.

Concurrency

• CONCURRENT_WRITES: .agents/ defaults to single-writer per session.
• Genuine multi-agent runs isolate each session under .agents/tmp/sessions/<session-id>/ and reconcile during the next maintenance pass rather than writing to shared memory/, rules/, workflows/, or skills/ concurrently.

Maintenance hygiene

• New trigger: aggregate .agents/memory/ content > ~3,000 lines.
• New cleanup actions: Age outcomes and Re-review harmful demotions.
• "Don't add noisy or one-off notes to .agents/" elevated from Maintenance Cadence to Hard Constraints (Must not do).

Compatibility

• Backward compatible with v1.3.0 installs: existing .agents/ layouts continue to work; new schema fields and thresholds activate as the next outcomes are recorded.
• Fully mirrored in AGENTS.zh-CN.md.

Diff

• Compare v1.3.0...v1.4.0

AgentGo v1.3.0

AgentGo v1.3.0

What's Changed

• Bumped the AGENTS.md template marker to v1.3.0.
• Added a lifecycle-based self-evolution model: fitness signals, memory lifecycle, capability lifecycle, outcome ledger, experiment isolation, and human feedback signals.
• Expanded .agents/ structure with memory/outcomes.md, experiments/, tmp/, reports, and runtime-supported skills guidance.
• Added periodic health-check behavior for memory freshness, structure drift, temporary files, workflow/skill promotion, and optional health reports.
• Added boundaries for high-confidence proactive suggestions: suggest optional improvements with evidence, but do not execute them without user request.
• Clarified template update policy: preserve the installed language, do not silently replace AGENTS.md on a timer, and use SemVer-like versioning.

Install

English:

curl -fsSL https://raw.githubusercontent.com/yeasy/agentgo/v1.3.0/AGENTS.md -o AGENTS.md

Simplified Chinese:

curl -fsSL https://raw.githubusercontent.com/yeasy/agentgo/v1.3.0/AGENTS.zh-CN.md -o AGENTS.md

AgentGo 1.0.0

Initial public release of AgentGo.

Highlights:

• Stable AGENTS.md protocol for reusable agent operating rules.
• Adaptive .agents/ project layer for durable memory, rules, workflows, and optional runtime-supported skills.
• English and Simplified Chinese documentation.
• Safety model for source priority, prompt-injection defense, high-risk actions, and evidence-based completion.