Task budgets and parallel agents reshape AI developer workflows

This week's developer-facing AI releases cluster around one theme: giving autonomous coding agents longer, more predictable runs. Anthropic shipped Claude Opus 4.7 alongside new API primitives for bounding and steering agent loops, OpenAI pushed Codex toward running multiple agents in parallel, and GitHub Copilot moved its autonomous Autopilot mode into public preview. Several of the changes land inside the CLIs developers already reach for — new slash commands in Claude Code and auto model selection in the Copilot CLI — rather than in chat-only surfaces. The shared direction is finer control over cost, concurrency, and review surface for agent runs.

Task budgets for Claude agent loops

Anthropic released Claude Opus 4.7 on April 16, 2026, and introduced task budgets in public beta alongside it (Anthropic release notes).

Task budgets let an application cap the total tokens Claude can spend on a single agentic turn — including thinking, tool calls, tool results, and output — by setting a taskBudget value and opting in with the task-budgets-2026-03-13 beta header (Anthropic task budgets documentation). The model receives a running countdown of remaining tokens and uses it to prioritize work and finish gracefully, which is distinct from max_tokens, the hard per-request ceiling. The same release adds the advisor tool in beta under the advisor-tool-2026-03-01 header, which pairs a faster executor model with a higher-intelligence advisor that provides strategic guidance mid-generation (Anthropic advisor tool documentation). Claude Opus 4.7 is available on the Claude platform, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry, at the same pricing as Opus 4.6.

These primitives enable long-running code-modification agents to stay within a known token cost per run instead of relying on a step-count limit that does not correlate well with spend. A CI agent triggered on every failing build can be bounded to a fixed number of tokens per ticket, letting engineering teams forecast monthly infrastructure cost for autonomous bug fixers. The advisor tool pattern — a Sonnet executor at medium effort paired with an Opus advisor — approximates Sonnet-at-default-effort intelligence at lower cost on coding tasks, according to Anthropic's guidance.

Parallel agents in OpenAI Codex and the Agents SDK

OpenAI shipped a major Codex update alongside the next evolution of its Agents SDK this month, moving toward concurrent autonomous agents and a unified Responses API foundation (OpenAI developer platform release notes).

The Codex platform now supports multiple concurrent AI agents working in parallel on a single project, running asynchronously in isolated git worktrees to avoid merge conflicts, with developers able to monitor each agent's state and review each agent's diff (OpenAI Codex changelog). The Agents SDK update adds a model-native harness with native sandbox execution, configurable memory, standardized integrations, portable workspace support, and built-in snapshotting for durable agent runs. The Responses API is the recommended foundation for new agent development going forward, with the Assistants API scheduled for deprecation in mid-2026.

Parallel worktrees change how a team can distribute work across a backlog: several Codex agents can pick up independent issues simultaneously and submit separate diffs, so an engineer reviewing PRs in the morning finds work produced overnight across multiple tickets. Snapshotting and durable runs in the Agents SDK let an agent survive container restarts without losing its state, which supports overnight refactors or long-running evaluation pipelines that previously required custom checkpointing logic.

Claude Code workflow additions

Claude Code added /ultrareview, /less-permission-prompts, a session recap, a new /effort slider, and extended prompt caching controls in its April releases (Claude Code changelog).

/ultrareview runs a comprehensive code review in the cloud using parallel multi-agent analysis and critique, and /less-permission-prompts scans session transcripts for common read-only Bash and MCP calls and proposes a prioritized allowlist for .claude/settings.json. The /effort command without arguments now opens an interactive slider with arrow-key navigation between levels and Enter to confirm, and the release adds 1-hour and forced 5-minute prompt caching controls plus a session recap feature configurable in /config and invocable with /recap. An interactive Google Vertex AI setup wizard also shipped on the login screen under the 3rd-party platform option, covering GCP authentication, project and region configuration, credential verification, and model pinning.

The permission-prompt skill reduces repeated approval dialogs on read-only commands, which affects any team running Claude Code in agent mode against a large repository. The longer prompt caching windows let applications reuse large system prompts across sessions that span up to an hour without re-billing for cached tokens, and /ultrareview enables a cloud-side parallel review on a branch before pushing to CI.

GitHub Copilot Autopilot and auto model selection

GitHub shipped Autopilot for fully autonomous agent sessions in public preview and made Copilot auto model selection generally available in the Copilot CLI for all Copilot plans (GitHub Changelog, April 2026).

Autopilot runs an agent session without per-step approval, and Copilot CLI auto model selection lets Copilot pick the most efficient model per request, with additional CLI features including short session-ID prefixes, a --print-debug-info flag, and usage-limit warnings (GitHub Copilot CLI release notes). Claude Opus 4.7 is rolling out across Copilot, and the platform now supports data residency for US and EU regions so that all inference processing and associated data stay within the designated geography. Organizations can also selectively enable GitHub Copilot cloud agent access on a per-organization basis.

Autopilot supports scheduled code-modification jobs — for example, a nightly dependency-bump agent that opens one PR per upgraded package without human intervention on each step. CLI auto-selection lets terminal workflows run mixed-complexity tasks without a developer hard-coding a model per command, and regional data residency removes a common blocker for Copilot adoption in organizations with EU-only data handling requirements.

Closing

The week's developer-facing releases converge on workflows where autonomous agents run for minutes to hours: bounded by task budgets on Anthropic's API, parallelized across git worktrees in Codex, scheduled through Copilot Autopilot, and tuned per session through new Claude Code commands. The shared shape across vendors is finer control over cost, concurrency, and review surface for agent runs, rather than incremental gains in chat-style completion quality. Teams operating autonomous coding agents now have documented primitives for capping spend per run, distributing work across parallel worktrees, and keeping inference within regional boundaries.

Sources

• https://www.anthropic.com/news/claude-opus-4-7
• https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7
• https://platform.claude.com/docs/en/build-with-claude/task-budgets
• https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool
• https://platform.claude.com/docs/en/release-notes/overview
• https://code.claude.com/docs/en/changelog
• https://developers.openai.com/codex/changelog
• https://developers.openai.com/api/docs/changelog
• https://github.blog/changelog/month/04-2026/
• https://github.blog/changelog/2026-04-17-github-copilot-cli-now-supports-copilot-auto-model-selection/
• https://github.blog/changelog/2026-04-08-github-copilot-in-visual-studio-code-march-releases/