Within roughly ninety days, six companies shipped or fundamentally restructured a coding-agent harness — the orchestration layer that assembles context, calls tools, and turns a model's output into something that runs. Google expanded Antigravity into a full platform at I/O 2026 (May), retiring its own Gemini CLI in the process.[1] GitHub announced a shared harness now powering Copilot CLI, the GitHub Copilot app, code review, and SDK experiences (June), with Copilot CLI becoming the default even for JetBrains.[2] xAI shipped Grok Build in beta (May 25), open-sourced the harness itself (July 15), and paired it with Grok 4.6 (August 12).[3][4] Meta launched Muse Code in beta (August 5-6) — its first coding product from Meta Superintelligence Labs.[5] DeepSeek released Harness v0.1 in developer preview, MIT-licensed (August 13).[6] Microsoft, separately, confirmed on its July 29 earnings call it's folding chat, coding, and other tools into one 'Super App,' merging its consumer and enterprise Copilot apps on August 13.[7][8] One company bet the opposite way: Block's Buzz (July 21) wraps existing harnesses — Goose, Codex, Claude Code — rather than building a proprietary one.[9] This library has independently tracked what happens once this category ships: UC-198's 1.7x-more-bugs, tripling-CVE, $4B remediation wave; UC-082's Amazon 6.3-million-order outage; UC-042's Antigravity-specific context-amnesia reports; UC-069's zero-click Copilot vulnerability; and UC-253's own finding that 'the attack surface is the harness, not the weights.'[10][11][12][13][14]
The compression is the story. Google expanded Antigravity into a full platform at I/O 2026 in May — a standalone desktop app, a CLI that formally retired Google's own Gemini CLI, and an SDK for self-hosting the same harness.[1] GitHub announced in June that one shared harness now powers Copilot CLI, the GitHub Copilot app, code review, and SDK-based experiences — Copilot CLI became the default even inside the JetBrains integration.[2] xAI shipped Grok Build in beta May 25, open-sourced the harness itself (the Rust agent loop, TUI, and tool layer) on July 15, then paired it with the new Grok 4.6 model on August 12.[3][4] Meta launched Muse Code in beta August 5-6 — the first coding-specific product out of Meta Superintelligence Labs, built on the new Muse Spark 1.2 model, which Mark Zuckerberg announced would be open-sourced days later.[5] DeepSeek released Harness v0.1 in developer preview, MIT-licensed, on August 13 — built so every component (model, tools, sessions, sandbox, UI) is a replaceable plugin.[6]
Microsoft moved on a different axis entirely. CEO Satya Nadella confirmed on the July 29 earnings call that Microsoft is folding chat, coding, its Cowork research tool, and Autopilot into a single 'Super App.' On August 13 — the same day DeepSeek shipped its harness — Microsoft actually merged its consumer and enterprise Copilot apps into one, cutting several features (Group Chats, AI podcasts, Copilot Labs, Deep Research) in the process.[7][8]
One company took the opposite bet entirely. Block, led by Jack Dorsey, launched Buzz on July 21 — not a new harness, but a workspace that wraps existing ones. Its CLI ships integrations for Goose, Codex, and Claude Code rather than a proprietary alternative; asked directly on X whether Buzz supports 'bring your own harness,' Dorsey's reply was a one-word 'yes.'[9] Every other company on this list bet on owning the layer. Block bet on being the layer everyone else's harness plugs into.
This library has independently tracked what happens once this exact category of tool ships. UC-198 found AI coding tools producing code with 1.7x more bugs, tripling CVEs, and a $4B remediation wave.[10] UC-082 found AI coding agents destroying production systems faster than guardrails can be built — Amazon lost 6.3 million orders in one AI-related outage, one of 10+ destruction events across 6 tools in 16 months.[11] UC-042 named Antigravity specifically: users reporting context amnesia, GitClear measuring 4x code duplication.[12] UC-069 covered Microsoft's own zero-click Copilot vulnerability, exfiltrating enterprise data with no user action required.[13] UC-253's own framing states the pattern almost exactly: 'the attack surface is the harness, not the weights' — EchoLeak exploited Copilot's pipeline specifically, not any model.[14] None of this proves the newest entrants are unsafe — there's no incident tied to Grok Build, Muse Code, or DeepSeek Harness yet. What's real and dated is the pattern, and two of the specific products already implicated in it — Copilot and Antigravity — are the same two that just expanded fastest.
How six companies' harness moves compressed into ninety days — and where this library had already been tracking the cost.
Full platform at I/O 2026: desktop app, CLI, and SDK for self-hosting the same harness.
Move 1Beta launch, then the Rust agent loop, TUI, and tool layer open-sourced a month and a half later.
Move 2CLI, app, code review, and SDK all share one harness — CLI becomes the default even for JetBrains.
Move 3Buzz wraps Goose/Codex/Claude Code instead of building a rival. Nadella confirms coding folds into one Microsoft app.
Move 4Muse Code beta, Grok 4.6 ships, DeepSeek Harness v0.1 and Microsoft's Copilot app merger land the same day.
Move 5The attack surface is the harness, not the weights. — This library's own UC-253, on why EchoLeak exploited Copilot's pipeline rather than any model
| Dimension | Evidence |
|---|---|
| Operational (D6) Origin · 88 | Six companies shipped or restructured a coding-agent harness within roughly ninety days, each independently confirmed against the company's own announcement.[1][2][3][5][6][7][9]The 90-Day Sprint |
| Quality (D5) L1 · 85 | This library's own UC-198, UC-042, and UC-253 already found the harness layer specifically responsible for bugs, CVEs, code duplication, and security exploits.[10][12][14]Already Documented Harm |
| Customer (D1) L1 · 72 | UC-082 disclosed Amazon lost 6.3 million orders in a single AI-related outage — the sharpest customer-facing harm figure in the evidence set.[11]Amazon's Disclosed Figure |
| Revenue (D3) L2 · 68 | UC-198's $4B bug-remediation wave and UC-069's $190M security-response funding round attach real cost to the pattern.[10][13]The Dollar Figures |
The cascade originates in D6 — Operational — because the lever is a real, dated wave of product and platform decisions: six companies shipping or restructuring a coding-agent harness within roughly ninety days. From D6 it cascades to D5 (Quality — this library's own prior findings on what this exact category of tool already causes: UC-198's bug/CVE data, UC-042's Antigravity-specific context-amnesia and duplication findings, UC-253's harness-as-attack-surface framing) and D1 (Customer — UC-082's disclosed Amazon figure, 6.3 million lost orders in a single AI-related outage, the sharpest customer-facing harm in the evidence set). It reaches D3 (Revenue — UC-198's $4B remediation wave and UC-069's $190M security-response funding round, real dollar figures attached to the fallout). D2 and D4 are deliberately left unscored — no disclosed workforce or regulatory figure ties specifically to this 90-day window.
-- UC-311: Six Companies, Ninety Days: 6D Diagnostic Cascade
-- Six companies shipped or restructured a coding-agent harness within ~90 days (May-Aug 2026): Google Antigravity (I/O May, retired Gemini CLI), GitHub Copilot shared harness (June), xAI Grok Build (beta May 25, open-sourced Jul 15, Grok 4.6 Aug 12), Meta Muse Code (beta Aug 5-6, first MSL coding product), DeepSeek Harness v0.1 (MIT, Aug 13), Microsoft Super App consolidation (confirmed Jul 29, Copilot apps merged Aug 13). Block's Buzz (Jul 21) bet the opposite way - wraps Goose/Codex/Claude Code rather than building a proprietary harness. This library independently tracked the risk side across UC-198 (bugs/CVEs/$4B remediation), UC-082 (Amazon 6.3M-order outage), UC-042 (Antigravity context amnesia), UC-069 (zero-click Copilot vuln), UC-253 (harness as attack surface, EchoLeak/Copilot).
FORAGE six_companies_ninety_days
WHERE harness_race_timeline_confirmed = true
AND library_risk_evidence_already_tracked = true
AND block_buzz_interoperability_bet_confirmed = true
ACROSS D6, D5, D1, D3
DEPTH 3
SURFACE six_companies_ninety_days
DIVE INTO harness_race_accelerates_amid_known_risk
WHEN new_harness_launches_dated_confirmed = true
AND prior_library_cases_already_document_harm = true
TRACE ai_developer_tooling_risk_cascade
EMIT harness_race_signal
DRIFT six_companies_ninety_days
METHODOLOGY 90
PERFORMANCE 40
FETCH six_companies_ninety_days
THRESHOLD 1000
ON MONITOR CHIRP high 'Six companies shipped or restructured a coding-agent harness within roughly 90 days: Google's Antigravity (I/O 2026, May, retiring Gemini CLI), GitHub's shared Copilot harness (June, default even for JetBrains), xAI's Grok Build (beta May 25, harness open-sourced Jul 15, paired with Grok 4.6 Aug 12), Meta's Muse Code (beta Aug 5-6, first Meta Superintelligence Labs coding product, Muse Spark 1.2 to be open-sourced), DeepSeek's Harness v0.1 (MIT-licensed developer preview, Aug 13), and Microsoft's Super App consolidation (confirmed on Jul 29 earnings call, Copilot apps merged Aug 13, same day as DeepSeek's release). Block's Buzz (Jul 21) bet the opposite way, wrapping Goose/Codex/Claude Code rather than building a proprietary harness - Jack Dorsey confirmed 'bring your own harness' support. This library independently found, across five prior cases, that the harness layer is where these tools actually break: UC-198 (1.7x more bugs, tripling CVEs, $4B remediation wave), UC-082 (AI coding agents destroying production faster than guardrails can be built, Amazon lost 6.3M orders in one outage), UC-042 (Antigravity users reporting context amnesia, GitClear found 4x code duplication), UC-069 (Microsoft's own zero-click Copilot vulnerability), UC-253 ('the attack surface is the harness, not the weights' - EchoLeak exploited Copilot's pipeline specifically).'
SURFACE analysis AS jsonRuntime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905332
Google, GitHub, xAI, Meta, DeepSeek, and Microsoft all chose to build or consolidate around a proprietary harness within the same 90 days.[1][2][3][5][6][7]
Block's Buzz wraps other companies' harnesses instead of building one — a bet on being infrastructure for the race, not a combatant in it.[9]
Five prior, independently-built cases document bugs, outages, and security breaches tracing to the harness layer specifically — not the underlying models.[10][11][12][13][14]
Copilot (UC-069, UC-253) and Antigravity (UC-042) are both named in this library's prior risk findings — and both just shipped major harness expansions.[1][2][12][13][14]
Each company's own announcement anchors the timeline; the risk evidence is drawn directly from this library's own previously-audited cases, cited as internal cross-references rather than restated as new claims.
Google, GitHub, xAI, Meta, DeepSeek, and Microsoft all shipped or restructured a coding-agent harness within ninety days. One company, Block, bet the opposite way — on being the layer everyone else's harness plugs into.