Agent Techniques Weekly
Big Read
OpenAI reported taking its ARC-AGI-3 public-set score from 13.3% to 38.3% by changing two Responses API harness settings on the same model: retain private reasoning between tool calls, and replace rolling truncation at 175k characters with context compaction. It also reported roughly 6× fewer output tokens. That is the week's central technique result — and it is vendor-measured on one benchmark by the company selling the API, with no independent replication yet. Discount it accordingly. The practical instruction does not need the replication to hold: before you attribute a performance gap to model quality, re-run the evaluation with identical weights and a different memory policy. The transferable pattern is Retained-Reasoning Compaction. Truncation forces the model to re-derive plans every turn once the window rolls; discarding private chain-of-thought between tool calls throws away hypotheses the next action needed. Compaction keeps learned environment state inside the window as a summary rather than a hard cut. Measure both task success and output tokens. If your harness still hard-truncates long trajectories, you are measuring the truncation policy as much as the model. The rest of the week made the same point from the connector and routing side. MCP specification 2026-07-28 went final as a normative, stateless request/response core — no initialize handshake, no session ID, method and name in headers, MRTR for mid-call elicitation, cacheable list results, and a twelve-month deprecation floor as Dynamic Client Registration yields to client metadata documents. Anthropic shipped a 950+ connector directory with IdP-provisioned enterprise connectors and research-preview MCP tunnels for private-network servers without inbound exposure. GitHub took Agent skills plus read-only MCP to general availability in Copilot code review on Jul 29. Snowflake positioned Cortex AI Gateway as a data-plane control plane for MCP and third-party agents, with many gateway features still private preview. Put together: fix the memory policy, then fix the connector contract, then encode task-class routing as harness policy. Microsoft's MDASH with MAI-Cyber-1-Flash handling roughly 90% of tasks before escalating to frontier models is the same pattern as Luna-default / Terra-mid / Sol-exception economics on GPT-5.6 fleets — specialist first, expensive generalist last. Independent audited workflow ROI with baseline and method still did not clear the bar this week. Treat CyberGym, ARC-AGI, and GPU-cost claims as vendor instrumentation until someone else reproduces them.
Technique of the Week
In long-horizon tool loops, keep private chain-of-thought between actions and summarize old context instead of hard-truncating it, so the agent preserves plans and environment state without paying a full re-derivation tax each turn.
OpenAI's vendor-reported ARC-AGI-3 jump (13.3% → 38.3%, ~6× fewer output tokens) came from harness memory settings on unchanged weights. That does not prove every workload will triple, but it does prove memory policy is a first-class evaluation variable. Operators who only swap models are confounding scaffold effects with intelligence. Fix the retention and compaction contract, instrument success and tokens, then decide whether a weight change is even warranted.
A multi-step tool loop approaches the context ceiling, or successive tool calls discard private reasoning between turns. The failure mode is repeated re-planning and rising output tokens without commensurate task progress.
Keep private reasoning items across tool calls rather than stripping them at each hop. The next action should inherit hypotheses, constraints, and partial plans already formed — not reconstruct them from truncated transcripts.
When the window fills, replace rolling hard truncation (OpenAI's prior official harness cut at 175k characters) with a summary compaction of older turns that preserves environment state, decisions taken, and open questions.
Use an API surface that can thread prior response state and compaction (OpenAI documents this via the Responses API with retained reasoning). Homegrown harnesses should store reasoning items and emit an explicit compaction artifact the model can read.
Score the same task suite under both memory policies with identical weights. Track task success, output tokens, retries, and verifier pass rate. A score move without a token move, or the reverse, is a scaffold signal — not a model ranking.
If compaction summaries lose critical state, split the trajectory into checkpointed stages with human review rather than silently truncating. Do not escalate to a larger model until the memory-policy A/B has been run.
OpenAI / automate
Vendor-reported ARC-AGI-3 public-set score rose from 13.3% to 38.3% with retained reasoning plus context compaction on the same model, replacing rolling truncation at 175k characters, with roughly 6× fewer output tokens.
Re-run agent evaluations with identical weights and a different memory policy before attributing a gap to the model. Treat the triple as a vendor-measured scaffold effect until independent parties reproduce it.
Microsoft / automate
A compact cyber specialist inside MDASH handles roughly 90% of vulnerability find/validate/remediate tasks, escalating harder cases to GPT-5.4-class models; Perception adds continuous monitor/patch agent teams.
Encode task-class routing — cheap specialist to expensive generalist — as harness policy with route logs, not as chat defaults. Demand evidence of when Flash versus frontier models actually ran.
GitHub / build
Agent skills via SKILL.md under .github/skills and MCP into code review reached GA; all MCP calls are read-only, with GitHub and Playwright MCP on by default, covering Pro through Enterprise.
Ship org coding standards as versioned skills and wire issue/docs MCP with read-only as the adoption gate. Prefer the same skill pack across cloud agent, CLI, and review rather than one-off prompts.
GitHub / xAI / build
Grok 4.5 is available in Copilot with up to 500k context, text-plus-image input, and reasoning-effort tiers; Business and Enterprise leave it off by default until an admin enables it.
Add Grok as an optional coding route only after admin policy and data-retention review. Long context is not a substitute for compaction discipline on multi-hour agent jobs.
Google DeepMind / automate
Embodied-reasoning orchestrator with live-video progress tracking, tool calling, VLA handoff, multi-robot collaboration, and the ASIMOV-Agentic refusal/escalation safety benchmark.
Treat physical agents as the same pattern as software agents — planner plus actuator plus safety stop — and demand ASIMOV-Agentic and human-proximity evidence before production pilots.
Alibaba / Qwen / cowork
Prior-session @ mentions inject read-only compressed summaries without resume or fork; completed background agents stay resident and recoverable across parent-session restore; /learn distills videos into reusable SKILL.md packs.
Session memory is becoming a first-class UX primitive. Prefer inject-summary over resume/fork when you need context without hijacking the current thread, and version learned skills like code.
OpenAI / automate
GPT-5.6 pricing and tiering make Luna viable for high-volume tool-using workflows, Terra for mid-complexity paths, and Sol Fast for latency-sensitive exceptions rather than a single frontier default.
Redesign agent graphs around Luna default, Terra mid, Sol exception. Pair the economic ladder with task-class routing so the expensive node is an escalation, not the idle default.
connector / Model Context Protocol
Normative final: stateless request/response core with no initialize or session ID; Mcp-Method and Mcp-Name headers; MRTR elicitation; cacheable list results; Tasks, MCP Apps, and Enterprise Managed Authorization as extensions; DCR deprecated toward CIMD with a twelve-month floor; Tier-1 SDKs ship same day.
Inventory every MCP server for hidden session assumptions, plan gateway routing on headers, prefer EMA for enterprise SSO, and treat Dynamic Client Registration as transitional debt under RFC 9207 issuer hardening.
connector / Anthropic
950+ connectors, IdP-provisioned enterprise connectors, research-preview MCP tunnels for private-network servers without inbound exposure, and connector performance dashboards.
Prefer IdP-provisioned connectors over user-owned OAuth sprawl. Use tunnels for private tools only in preview with explicit network policy; production still needs a gateway audit trail.
harness / Snowflake
Central policy, audit, and cost control plane for first- and third-party agents over models, tools, and MCP servers; Agent Identity GA; many gateway and partner-identity features remain private preview.
Put MCP behind a data-plane gateway before enabling IDE agents against warehouse tools. Treat preview controls as non-contractual until GA dates are in writing.
skill / GitHub
Org standards ship as SKILL.md under .github/skills; MCP context enters reviews as read-only calls, with cloud-agent MCP configs reusable across surfaces.
One versioned skill pack can serve cloud agent, CLI, and review. Read-only MCP is the right default for any automated reviewer that can see production systems of record.
Evidence: benchmark
Vendor-reported public-set score rose from 13.3% (official harness with rolling truncation) to 38.3% with retained reasoning plus compaction on GPT-5.6 Sol, with roughly 6× fewer output tokens.
This is the week's strongest technique evidence and still grade-discounted: same company, same API surface, one benchmark, no independent reproduction. Use it as a mandate to A/B memory policy under pinned weights — not as proof that every agent will triple overnight.
Evidence: vendor_claim
Vendor claims 96% CyberGym (+12 points vs Mythos) and roughly 50% cost versus the prior GPT-5.4 + 5.4 mini + 5.3 Codex stack by handling about 90% of tasks on MAI-Cyber-1-Flash.
Useful as a routing-pattern illustration, not as audited ROI. Require route logs and an external or red-team corpus before treating CyberGym percentages as procurement truth.
Evidence: vendor_claim
Vendor reports up to 84% lower GPU cost after switching production image generation to MAI-Image-2.5-Flash versus GPT-Image-2, plus a +26% OneDrive image-edit save rate on the same Flash default.
Production switch claims are directionally useful for completed-task cost reviews, but they are vendor instrumentation without published methodology detail. Rebid high-volume image paths on measured GPU/token cost in your own tenant.
MCP 2026-07-28 plus Claude IdP-provisioned connectors push enterprise SSO and EMA as the preferred path. Ban agent-inherits-full-user-token patterns; prefer task-scoped short-lived credentials as Snowflake's agent-identity model describes.
GitHub's read-only MCP default for code review is the right posture for automated reviewers. Research-preview MCP tunnels that reach private networks without inbound exposure still need explicit allowlists and gateway logging before production.
Stateless MCP moves routing onto Mcp-Method and Mcp-Name headers — good for gateways, useless without capturing those headers, effective model, memory policy (retain/compact), and skill-manifest version in the run log.
Retained reasoning and compaction change what the model 'knows' mid-run. Re-verify after compaction events the same way W30 required re-verification after model failover; summary loss is a silent failure mode.
Pair Luna/Terra/Sol (or Flash/specialist/frontier) ladders with measured output tokens under compaction. A cheaper model with truncate-and-retry can outspend a mid tier that retains and compacts cleanly.
GitHub's default model enablement for Business/Enterprise (effective Aug 26) turns new GA models on unless the org opts out. Set policy to disabled before that date if manual model approval is required; use enterprise-teams targeting (preview) for frontier experimentation without org-wide exposure.
Expected outcome: You leave with a side-by-side score and token delta under identical weights, plus a written rule: no model comparison ships without a declared memory policy. Expect at least one workflow where truncation was silently taxing performance.
Next 30 days
The week’s headline number stays vendor-grade until a third party reproduces retained-reasoning plus compaction under published harness settings.
By Aug 26
New GA models turn on for Business/Enterprise unless policy is set to disabled — operators who require manual approval must act before the effective date.
Aug 2026
Stateless 2026-07-28 plus the CIMD deprecation floor will expose which production connectors still assume initialize, session IDs, or Dynamic Client Registration.
Next 14 days
Many cost, routing, and partner-identity controls remain private preview; contractual GA timing decides whether the gateway is a production control plane or a pilot.