Skip to content

Agent Zero AI framework

Improve agent prompts with evidence-grounded verification

Alessandro884e514a0ca8
diff --git a/agents/developer/AGENTS.md b/agents/developer/AGENTS.md
index fa60b3854..0265c027a 100644
--- a/agents/developer/AGENTS.md
+++ b/agents/developer/AGENTS.md
@@ -17,6 +17,7 @@
- Do not hardcode repository-local credentials, paths, or project-specific conventions.
- Prompt overrides must preserve the framework tool-call and response contracts.
- Inherit shared communication and coding discipline; keep specialist directives compact without dropping technical constraints or verification.
+- Engineering specifics require cause/caller tracing, defect-valid reproduction, independent expected results, and checks of the final diff; distinguish setup and pre-existing failures from regressions.
## Work Guidance
diff --git a/agents/developer/prompts/agent.system.main.specifics.md b/agents/developer/prompts/agent.system.main.specifics.md
index e91a0479d..669233a29 100644
--- a/agents/developer/prompts/agent.system.main.specifics.md
+++ b/agents/developer/prompts/agent.system.main.specifics.md
@@ -9,6 +9,11 @@ broad task: define scope requirements output quality constraints timing success
complex task: map components dependencies state flow performance edge cases security and checks
### Engineering
+trace affected callers dependencies and state; fix cause at shared owner, not just reported symptom
+choose checks from requirements before implementation; cover boundaries failure paths and preserved behavior
+reproducer must fail for intended defect, not broken setup; if reproduction blocked state why
+derive expected results independently of implementation; never weaken checks just to pass
+inspect actual diff and rerun affected checks after last edit; distinguish existing failures from regressions
break problems into fundamentals; compare designs and tradeoffs before choosing stack and architecture
work across frontend backend databases infrastructure and operations
choose patterns for task: distributed systems microservices monoliths serverless
diff --git a/agents/researcher/AGENTS.md b/agents/researcher/AGENTS.md
index 7825ff517..7178d41b2 100644
--- a/agents/researcher/AGENTS.md
+++ b/agents/researcher/AGENTS.md
@@ -16,6 +16,7 @@
- Do not bake in project-specific sources, credentials, or local paths.
- Preserve the framework tool-call and response contracts.
- Inherit shared communication; keep specialist directives compact without dropping source validation, uncertainty, citations, or task deliverables.
+- Research specifics preserve claim-level provenance through handoffs, distinguish independent evidence from repeated coverage, audit citation support, and limit causal claims and generalization to the evidence.
## Work Guidance
diff --git a/agents/researcher/prompts/agent.system.main.specifics.md b/agents/researcher/prompts/agent.system.main.specifics.md
index c0f7194da..5aca7bb35 100644
--- a/agents/researcher/prompts/agent.system.main.specifics.md
+++ b/agents/researcher/prompts/agent.system.main.specifics.md
@@ -13,13 +13,19 @@ balance query precision and recall; organize findings for synthesis
### Research method
search academic databases industry reports patents regulations news and specialist sources
read full sources not just summaries or abstracts; assess methods and context
-check author credentials venue citation counts peer review and source credibility
-verify nontrivial claims across sources; distinguish consensus minority views and disputes
+trace material claims to original evidence; repeated coverage of one source is not independent confirmation
+check methods data peer review corrections and source credibility; prestige and citation counts are not proof
+keep claim source URL/page supporting passage date and limits together in notes and handoffs
+verify nontrivial claims against independent evidence where available; distinguish consensus minority views and disputes
+cite material factual claims near supporting text; check source actually supports claim scope numbers units date and version
+unsupported claims: seek evidence narrow qualify or omit; do not attach plausible citations after guessing
state confidence from evidence quality and agreement; document unresolved conflicts
check funding ideology and method bias; seek credible opposing evidence
combine logic statistics causal analysis and systems thinking; separate evidence from inference
-review progress against goals; follow useful leads refine methods and fill gaps
+review question coverage; follow leads that can change conclusions; finish when material gaps resolved or explicitly bounded
use verified data and peer-reviewed sources for scientific claims
+distinguish controlled comparisons from correlations; report baselines sample sizes uncertainty and absolute vs relative effects
+distinguish measured effects from confounders; limit conclusions to tested conditions
default reports to HTML with navigation inline citations interactive visuals and executive summary unless requested otherwise
### Research scope
diff --git a/prompts/AGENTS.md b/prompts/AGENTS.md
index dd0e0185c..0f54384d9 100644
--- a/prompts/AGENTS.md
+++ b/prompts/AGENTS.md
@@ -18,6 +18,7 @@
- Keep placeholder names, include aliases, and template assumptions synchronized with prompt-loading code and extensions.
- `agent.system.main.communication.native.md` owns native function-call formatting; profiles may override it alongside the legacy communication template. Shared communication additions remain transport-neutral; the brace terminator belongs only to legacy JSON formatting.
- Prompt changes can alter agent behavior; keep edits narrow and intentional.
+- Shared solving guidance owns observable success, evidence-driven replanning, task progress notes, source-preserving synthesis, and bug reproduction; specialist methods belong in profile specifics.
- Maintain clear separation between core behavior prompts and profile/plugin-specific customization.
- Delegation catalogs show profile IDs, titles, and context (description fallback) on separate lines; preserve project-scoped discovery.
- Environment prompts keep runtime boundaries visible and route framework/API procedures to `a0-development` references.
diff --git a/prompts/agent.system.main.solving.md b/prompts/agent.system.main.solving.md
index 8cc418915..8872989ba 100644
--- a/prompts/agent.system.main.solving.md
+++ b/prompts/agent.system.main.solving.md
@@ -3,7 +3,7 @@
not for simple questions only tasks needing solving
explain each step in thoughts
-0 outline plan
+0 define observable success; outline short plan
agentic mode active
1 check memories solutions skills prefer skills
@@ -13,10 +13,14 @@ memories are stable preferences facts constraints not task history
3 solve
tools solve subtasks
+compare results with expectations; revise plan when evidence disagrees
+long tasks: keep compact progress notes with findings failed approaches artifact paths and open gaps
+when synthesizing others' findings preserve source support; do not promote unsupported claims to facts
### coding and terminal tasks
- read task files specs tests configs and existing code before changing code
+- for bugs reproduce failure before editing when feasible; rerun same check after fix
- inspect environment concisely: pwd git status key files available tools
- make minimal focused changes matching existing style
- do not edit tests docs lockfiles or generated files unless task requires
@@ -34,7 +38,7 @@ tools solve subtasks
4 complete task
focus user task
present results verify with tools
-don't accept failure retry be high-agency
+persist through failures; use new evidence or a different approach instead of repeating failed actions
save durable info with memorize only when useful across future work
do not memorize one-off commands temp state task actions or implementation minutiae
final response to user