October 2026 Summaries
10 posts from Qodo
Filter
Month:
Year:
Post Summaries
Back to Blog
Qodo has introduced a capability that follows links from pull request tickets to external specifications and evaluates code changes against the requirements found there, helping reviewers detect gaps that standard diff review may miss. The feature is intended for spec-driven development workflows, where AI coding agents may implement ticket summaries correctly while overlooking constraints, edge cases, or security requirements documented in linked materials. In an example involving a bulk customer-record update, Qodo follows a Jira link to a Confluence API specification, identifies a missing account-level authorization check, and flags the resulting security risk. Supported specification sources include Asana, Confluence, Notion, ServiceNow, and JSM Assets, allowing teams to keep detailed requirements in their existing documentation systems while bringing that context into pull request review.
Oct 29, 2026
706 words in the original blog post.
Qodo improved its AI code-review recommendation agent, which predicts whether a specific team will act on a finding based on its historical pull-request behavior rather than technical correctness alone. To reduce inconsistent recommendations, the company adapted the GEPA prompt-optimization approach by comparing two agent runs per example, generating structured reflections on differences in evidence and decisions, and aggregating those reflections across a training set of roughly 400 PRs. Beginning with a minimal 43-word prompt, Qodo selected a revised instruction set that improved offline consistency and was then tested in a 50/50 production A/B experiment covering 27,582 review runs across about 21,000 PRs. The updated prompt reduced the share of implemented findings labeled “no” from 7.83% to 4.46%, cut average agent cost from $0.24 to $0.12 per run, and lowered latency from 40 to 30 seconds while maintaining comparable precision. Qodo attributed the results to clearer guidance on using historical evidence, avoiding redundant searches, managing uncertainty, and distinguishing team-response prediction from code-quality assessment, while emphasizing the value of minimal prompts, trace-based evaluation, and repeated measurements of agent variability.
Oct 22, 2026
1,156 words in the original blog post.
AI coding agents can pass local tests while overlooking broader commitments such as cross-repository client dependencies, prior review rationale, and support policies, making context-aware engineering decisions essential. Using the example of removing a deprecated API field still consumed by supported mobile clients, the discussion argues that agents need inspectable evidence from codebases, pull-request history, policies, and local worktrees to identify compatibility constraints and disclose remaining uncertainty. Agent memory can preserve prior reasoning, but it should record the evidence, scope, and conditions under which a decision can be reconsidered rather than treating past conclusions as permanent rules. Codebase intelligence, memory, standards, procedural skills, and execution harnesses serve complementary roles in bringing relevant knowledge into tasks and learning from review outcomes. Such systems should be evaluated with scenarios where the correct recommendation changes according to client support, migration status, relevance of a constraint, or gaps in available evidence, with emphasis on accurate citations, appropriate scope, and transparent reasoning.
Oct 20, 2026
1,258 words in the original blog post.
Qodo’s Code Quality and Governance platform v3.1 introduces workflow improvements aimed at reducing review noise, improving visibility, and simplifying administration. Its rule lifecycle system can now automatically mute rules whose findings are repeatedly ignored across sufficient merged pull requests, while the generally available analytics dashboard reports findings, accepted recommendations, and acceptance rates by factors such as severity, category, repository, and author. The release also adds a keyboard-accessible quick search tool for locating configuration settings and expands background activity reporting for repository synchronization, pull-request indexing, cross-repository discovery, and rule or skill imports. New status details identify indexing retries, incomplete Git installation setups, and direct links to affected installations or repositories, helping teams diagnose setup issues more efficiently.
Oct 19, 2026
689 words in the original blog post.
AI coding agents have increased development speed but also created a trust gap, as teams often must manually re-review agent-generated code, reducing productivity gains. A Qodo live session argued that code review should move earlier into the coding session rather than begin at the pull request, using organizational context such as cross-repository dependencies, engineering rules, and prior PR discussions to identify issues before changes are finalized. Qodo’s Agentic Toolbox integrates full AI review capabilities into coding agents, allowing developers to run iterative, detailed local reviews while their work and context remain active, including analysis across repositories not stored locally. The approach distinguishes between comprehensive findings for agents, which can handle high volumes of feedback, and lower-noise reviews for human PR reviewers. Under this model, developers can address both blockers and broader quality improvements before opening a PR, leaving pull requests as a final gate for prioritizing human attention, approving low-risk changes, and improving the system through feedback on accepted or dismissed findings.
Oct 17, 2026
1,180 words in the original blog post.
Qodo evaluated 13 AI models across 38 model and reasoning-level configurations for code review using a proprietary benchmark of 125 real pull requests containing 703 bugs that were later discovered and fixed. The assessment measured precision, recall, F1 score, backtracked-bug recall, cost per pull request, and median review time while holding the review setup constant. Claude Opus 5.5 achieved the highest listed F1 score at 56.8%, closely followed by Claude Sonnet 5.5, Claude Fable 5.1, and the open-weight GLM 5.3, while several other models formed a competitive second tier. Results indicate that higher reasoning settings can improve bug recall but often reduce precision, increase costs, and lengthen review times, making the optimal setting model- and priority-dependent. Cost varied by more than 100-fold across configurations, and open-weight models such as GLM 5.3, DeepSeek V4 Pro, and Kimi K3 performed competitively with proprietary alternatives. Qodo says it continuously reruns the benchmark for new releases and uses the findings to select production review models, offering customers provider flexibility and adaptive choices based on desired trade-offs among coverage, confidence, cost, and speed.
Oct 12, 2026
1,347 words in the original blog post.
Qodo describes Dark Factory, an autonomous yet human-supervised software engineering team designed to turn scoped requests into reviewed, merge-ready pull requests through collaboration among specialized AI agents. Built on BAND, a shared interaction layer supporting persistent identities, chat rooms, message routing, and cross-environment communication, the system addresses limitations of point-to-point protocols such as A2A, which enable delegation but do not inherently provide shared conversation history, membership, recovery, or collective context. Dark Factory combines a SWE agent that plans and implements changes, an independent Reviewer that validates work without editing it, and an externally hosted SRE agent that supplies privileged production evidence, while engineers can monitor, guide, or stop work through the same shared room. Deterministic tools manage consequential operations such as repository changes, pull request lifecycles, checkpoints, and review gates, preserving auditability alongside flexible agent reasoning. Qodo argues that this room-based model can allow independently deployed agents with different tools, permissions, and expertise to operate as a coordinated team, potentially extending to security, testing, release, and other specialized roles.
Oct 06, 2026
2,579 words in the original blog post.
An evaluation of TypeSafe’s Jev structured-judgment model alongside selected automated checks and Qodo code review found that Jev identified one additional required testing gap across 24 cases but no additional confirmed defects, while also raising unsupported concerns in three cases and escalating one verification request into an unsupported fix request. Its most useful contribution involved a controlled rollout case where submitted tests missed query-specific latency isolation despite healthy aggregate latency, whereas Qodo’s findings covered unrelated issues. The revised protocol compared checks alone, checks with Qodo, checks with Jev, and all combined, using independent executable probes to label adequate tests, missing test coverage, and seeded defects; Qodo found eight seeded defects and 15 of 16 required gaps, while workflows involving Jev requested all 16 gaps. However, differences in evidence access, reused reviews, previously seen cases, non-independent repeated passes, and unmeasured developer effort limited conclusions about comparative performance or net benefit. The study recorded substantial input usage and higher latency for query-focused packets, concluding that Jev merits further evaluation but should be constrained by named behavioral requirements, traceable evidence, preserved blockers, fresh cases, and measurement of both useful verification and unnecessary maintainer work.
Oct 05, 2026
1,006 words in the original blog post.
Qodo 3.0 is an AI code quality and governance platform designed to help enterprises manage software development as coding agents take on more work across the SDLC. It addresses challenges such as fragmented agent-generated pull requests, standards applied too late, limited visibility into complex dependencies, and deployment constraints in organizations using diverse Git providers or secure environments. Its new PR Triage feature groups related pull requests into cross-repository work packages, providing information on blast radius, review difficulty, waiting time, review order, priorities, and ownership. The Agentic Toolbox gives coding agents access to organizational rules, repository guidelines, codebase context, and iterative reviews before pull requests are opened, while the Software Map tracks service and repository dependencies to clarify a change’s potential impact. A new analytics dashboard provides leaders with quality findings, fix rates, unresolved issues, and drill-down access to underlying reviews. Qodo 3.0 also supports GitHub, GitLab, Bitbucket, Azure DevOps, and Gerrit, along with on-premises, air-gapped, and self-hosted model deployments, aiming to embed consistent quality controls throughout enterprise software factories.
Oct 01, 2026
1,335 words in the original blog post.
As enterprises adopt agentic software development lifecycles, coding generation is accelerating while code review, quality assurance, security, and governance increasingly become the main constraints. The author argues that effective software factories must combine generative workloads with independent quality systems that guide agents during development and assess interconnected groups of pull requests before release, rather than relying on a single final review checkpoint. Central to this approach is a “wisdom base” that captures organizational judgment from codebases, review histories, accepted rules, and engineering decisions, allowing quality practices to compound even as models and coding agents change. Qodo 3.0 is presented as a platform designed around this model, adding PR Triage for cross-repository work-package visibility, an Agentic Toolbox that supplies team-specific context to coding agents, multi-agent code review, PR Insights into learned review behavior, and a Software Map with quality metrics for leadership oversight, including support for enterprise environments such as Gerrit, on-premises deployments, air-gapped systems, and open-source models.
Oct 01, 2026
1,756 words in the original blog post.